Recent Surge in Public Interest in Transportation: Sentiment Analysis of Baidu Apollo Go Using Weibo Data

ArXiv: 2408.10088

🎯 Pitch

A single trending event in July 2024 triggered a massive 90% spike in online discussion about Baidu’s robotaxis, but it also flipped the sentiment script. While the discourse was largely positive up to that point, the viral moment unleashed a surge of anxiety about job displacement and safety that the initial tech enthusiasm had previously masked.

1. Executive Summary

This study analyzes public sentiment toward Baidu Apollo Go, a leading autonomous ride-hailing service in China, by applying a hybrid BERT model (fine-tuned bert-base-chinese) to 36,096 Weibo posts collected from January through July 2024. The analysis tracks public attitudes along two axes — temporal evolution (comparing the pre-surge period of January–June against the viral-trending period of July) and spatial distribution across Chinese provinces — revealing that 89.56% of all posts cluster in July alone and that sentiment, while predominantly positive overall, begins shifting toward negative after Apollo Go becomes a trending topic on July 11–12, with positive discourse centering on technology applications and personal experience and negative discourse centering on job displacement and safety concerns. The study further establishes that discussion intensity strongly correlates with provinces where Apollo Go operates (Guangdong, Beijing, Hubei dominating online volume), yet provinces without operational service exhibit sharply divergent attitudes — Xinjiang and Qinghai showing optimism while Tibet and Gansu express concerns about impacts on traditional taxi services — demonstrating that public acceptance of autonomous ride-hailing is geographically heterogeneous even within a single national market.

2. Context and Motivation

The Core Problem: Autonomous Ride-Hailing Is Deploying Faster Than We Understand Public Acceptance

The fundamental question this paper tackles is both timely and understudied: when a robotaxi service like Baidu Apollo Go undergoes rapid, large-scale commercial deployment in major cities, how does the general public actually feel about it — and how do those feelings vary across time and geography? This matters because, as the paper explicitly states, "autonomous ride-hailing services are rapidly growing globally, and the general public's acceptance will determine their success" (Section 5). The fate of these multi-billion-dollar technological deployments hinges not primarily on whether the vehicles work, but on whether the public wants them.

This gap is significant for several reasons that extend beyond academic curiosity:

• Policy and regulation depend on public sentiment: Transportation agencies and city governments must decide whether to permit, restrict, or subsidize autonomous ride-hailing operations. These decisions are politically mediated — public backlash can derail deployment timelines, while enthusiasm can accelerate them. Without systematic data on what drives positive and negative sentiment, policymakers are effectively operating blind.
• Market adoption requires understanding user concerns: Even if the technology is viable, potential riders who harbor safety fears or moral objections to labor displacement will not adopt the service. Service providers like Baidu need to understand which concerns are most salient in which demographic and geographic contexts to design effective outreach and mitigation strategies.
• The deployment is happening now, not hypothetically: Unlike earlier research that surveyed abstract attitudes toward hypothetical autonomous vehicles, Apollo Go is an operational service with real users encountering real vehicles on real streets. The public conversation captured in this study reflects responses to an actual technological transition, not speculative future scenarios.

The Gap: Real-Time, Large-Scale, Geographically Granular Sentiment Data Is Missing

Prior research on public attitudes toward autonomous vehicles falls into two broad categories, both of which the paper identifies as having significant limitations relative to studying an actively deployed commercial robotaxi service.

Survey-based studies are limited by hypothetical scenarios and small samples. The literature contains numerous survey studies probing public attitudes toward autonomous vehicles — Bansal et al. (2016) surveyed Austin residents about interest in new vehicle technologies; Montoro et al. (2019) studied perceived safety as a predictor of AV adoption intention among Spanish drivers; Roche-Cerasi (2019) assessed public acceptance of driverless shuttles in Norway. These studies typically work with sample sizes in the hundreds, rely on stated preferences rather than revealed behavior, and — critically — ask respondents to evaluate hypothetical technologies. As the paper notes, even before widespread deployment, the public held what it terms a "cautious attitude towards autonomous taxis" (Section 1), with technological advancement fostering trust while job displacement concerns and perceptions of dehumanization fostered negative sentiment (citing Tussyadiah et al., 2017). But these attitudes were measured in the abstract, before an actual widespread deployment put them to the test.

Existing social media analyses of AV sentiment have scope limitations. A handful of studies have used social media data to study autonomous vehicle sentiment — for example, Ding et al. (2021) analyzed Twitter feeds to understand how sentiments on autonomous vehicles are influenced by events; Jiang and Zheng (2021) used text mining of Twitter data to gauge public perceptions of driverless technologies in the United States. The paper acknowledges these as important precedents but identifies substantive gaps: most focus on Western markets (the U.S. or Europe), most analyze a general "autonomous vehicle" rather than a specific operational service, and few have the geographic granularity to disaggregate sentiment by province or region within a single country. Chen and Tomblin (2021), for instance, combined Reddit data, public deliberation transcripts, and surveys to study U.S. public opinion on AVs — a methodologically rich approach, but one tied to the American context and to discussions of AVs in general rather than an operational fleet.

The Chinese context presents unique dynamics. The paper situates itself within a specific ecosystem where prior work is scarce. Apollo Go is "China's first shared autonomous driving service platform operating in multiple cities" and is "scaling up its passenger testing operations" (Section 1). The Chinese social media landscape — centered on Weibo rather than Twitter/X — has distinct characteristics: Weibo posts often include IP-verified geographic location data (accurate to the province or municipality level) that Twitter does not provide, enabling the spatial analysis that forms a core contribution of this paper. Moreover, the Chinese public's concerns may differ from those documented in Western studies. The paper explicitly flags that in the U.S. and European AV discourse, "there is more discussion about personal privacy and data security" (Section 5), while in the Chinese Weibo conversations captured here, such concerns are notably absent — with the discourse instead centering on technology excitement, job displacement, and comparisons between autonomous and traditional taxi driver behavior. This suggests that findings from Western social media analyses may not transfer cleanly.

Where Prior Approaches Fall Short

The paper identifies four specific limitations in the existing research landscape that motivate its approach:

1. Survey research captures hypothetical attitudes, not reactions to lived experience. When Apollo Go began commercial operations in Wuhan in 2024 and the topic went viral on Chinese social media, the public was no longer speculating about a future technology — they were commenting on videos of actual vehicles navigating actual streets, sharing personal ride experiences, and reacting to real news about pricing and employment impacts. The paper argues that this shift from hypothetical to experiential discourse produces fundamentally different sentiment patterns than survey-based research can access. For example, the finding that some Weibo users express "dissatisfaction with the presence of safety operators in the vehicles, feeling that it undermined the autonomy of the service" (Section 4.1) is a nuanced experiential complaint that would not emerge in a hypothetical-scenario survey.

2. Most Western social media studies lack granular geographic data. The paper repeatedly emphasizes a distinctive advantage of Weibo data: "it contains the user's real IP address, offering a more precise measure of geographical data compared to X (Twitter), where users' locations are only inferred from their text" (Section 5). This is not a minor methodological detail — it enables the paper's core finding that attitudes toward Apollo Go vary dramatically by province, with Xinjiang showing deep red (positive) while neighboring Tibet shows deep blue (negative) (Section 4.2, Figures 9–10). A Twitter-based study that infers location from user bios or tweet text simply could not produce this level of geographic resolution, and could not test the claim that discussion intensity correlates with operational deployment regions.

3. Existing NLP approaches for Chinese social media sentiment analysis are underexplored for transportation topics. While the general BERT fine-tuning paradigm for Chinese text is well-established (the paper uses bert-base-chinese from Hugging Face), its application to the specific domain of autonomous ride-hailing sentiment is novel. The paper notes that social media text presents particular challenges — "metaphors, harsh data, and the coexistence of many possible meanings for individual words or phrases" (Section 1) — and that these challenges are amplified in the Chinese-language context, where word segmentation ambiguity and domain-specific neologisms (like colloquial names for Apollo Go) can degrade off-the-shelf model performance. The paper's manual annotation effort (2,697 labeled posts across four categories: positive, negative, neutral, and drop) and fine-tuning procedure on this domain-specific data represents a deliberate attempt to address this gap.

4. No study has captured the temporal dynamics of a robotaxi deployment "viral moment." Perhaps most distinctively, the paper documents what it frames as a natural experiment: Apollo Go was quietly operating from January through June 2024, generating modest online discussion (742 posts in May, 685 in June), when suddenly — likely triggered by the Apollo Day 2024 event on May 15 and subsequent media coverage — it became a trending topic in July, generating 19,597 posts in a single month (89.56% of the entire seven-month dataset). This is not a gradual ramp; it is a step-function change in public awareness. The paper uses this temporal structure to disaggregate "pre-trending" attitudes from "post-trending" attitudes, revealing that "negative comments began to rise after it became a hot topic on July 21" (Abstract) — a finding that would be invisible in a cross-sectional survey or a dataset aggregated across time.

How This Paper Positions Itself

The paper does not claim to develop new NLP architectures or to advance the theoretical foundations of sentiment analysis. Its positioning is entirely empirical and applied: it uses an established method (fine-tuned BERT for Chinese text sentiment classification) as an instrument to answer substantively important questions about public acceptance of autonomous ride-hailing during an active, large-scale commercial deployment.

This is evident in the three research questions the paper sets out to answer (Section 1):

1. How can we use natural language data to better understand new mobility services?
2. What is the general public's emotional reaction to the sudden appearance of the robotaxi?
3. What suggestions and insights can be informed to the public and transportation agencies?

These are not NLP research questions (they don't ask "how can we improve sentiment classification accuracy?") — they are transportation policy and urban planning questions for which NLP serves as the measurement tool. The paper positions its primary contribution as the findings that the NLP pipeline enables: the discovery that negative sentiment rises after trending events, that geographic sentiment is heterogeneous even within a single country, and that positive and negative discourse cluster around distinct thematic concerns (technology experience vs. labor displacement).

Theoretical framing: sentiment as a window into adoption barriers. The paper draws on established frameworks in transportation behavior research that treat "cognitive, emotional, and attitudinal factors" as significant influences on transportation mode choice (citing Parkany et al., 2004). By extracting these factors from large-scale social media discourse rather than small-scale surveys, the paper positions itself as providing a complementary measurement methodology that can overcome the scale and realism limitations of traditional attitude research. The paper is not contesting the validity of survey-based findings — it is augmenting them with a different data source that captures spontaneous, geographically distributed, temporally granular public discourse.

Domain-specific motivation: the Apollo Go moment. The paper anchors its contribution in the specific context of Apollo Go's 2024 expansion, arguing that this constitutes a "paradigm shift in the urban transportation landscape" (Section 1, citing Schaller, 2018 and Li et al., 2022). The claim is that the rapid deployment of robotaxi services in major Chinese cities represents a qualitatively new phase in autonomous vehicle adoption — moving from limited pilot programs to city-wide commercial operation — and that studying public sentiment during this transition is both urgent (because policy decisions are being made now) and uniquely informative (because the public is reacting to lived experience rather than hypothetical scenarios).

The international comparative dimension (implicit). While the paper does not frame itself primarily as a comparative study, it repeatedly gestures toward international comparisons: noting that findings are "similar with studies conducted in the U.S." (Section 5) for positive sentiment themes, but that privacy concerns — prominent in Western discourse — are absent from Chinese discussions, and that some negative Chinese comments focus on dissatisfaction with traditional taxi driver behavior (smoking, complaining about short trips) rather than solely on AVs themselves. This suggests a positioning where the Chinese case is neither assumed to be universal nor treated as sui generis — it is a specific deployment context whose findings can be compared with and contrasted against the growing international literature on AV public acceptance.

Methodological positioning: practical NLP for transportation research. The paper's choice to fine-tune an existing pretrained model rather than develop a novel architecture, and its transparent reporting of model accuracy (0.59), reflects a pragmatic orientation. The goal is not to achieve state-of-the-art sentiment classification performance but to demonstrate that off-the-shelf NLP tools, when fine-tuned on domain-specific labeled data, can produce meaningful insights for transportation researchers and policymakers who may lack deep NLP expertise. The release of the model on Hugging Face and the code repository on GitHub reinforces this positioning: the paper is providing a reproducible, accessible pipeline that others can apply to similar transportation-related social media analysis tasks.

Acknowledged limitations as boundary conditions. The paper is explicit about the scope of its contribution: the lack of user demographic information (age, income) limits the ability to explain why attitudes vary across groups; the seven-month observation window is shorter than typical in similar studies; and the data volume, while large by survey standards, is concentrated in a single platform (Weibo) and a single service (Apollo Go). These are not weaknesses of execution but deliberately acknowledged scope constraints that position the study as an initial empirical investigation into a rapidly evolving phenomenon, inviting follow-up work with longer time horizons, multi-platform data, and demographic controls.

3. Technical Approach

3.1 Reader Orientation

This paper is fundamentally a data-driven empirical analysis rather than a novel algorithm paper: the "system" being built is a practical NLP pipeline that takes raw Chinese-language social media posts mentioning Apollo Go, processes them through a series of cleaning and classification stages, and produces labeled sentiment data (positive, negative, neutral) that can be analyzed across temporal and spatial dimensions. The system solves the problem of how to measure public sentiment at scale — specifically, how to take 36,096 noisy Weibo posts and transform them into structured, analyzable sentiment labels that reveal when and where public attitudes toward autonomous ride-hailing are shifting.

3.2 Big-Picture Architecture (Diagram in Words)

The pipeline can be understood as a five-stage processing chain, with information flowing sequentially from raw data through cleaning, annotation, model training, classification, and finally analysis:

1. Data Collection — The Weibo API retrieves 36,096 posts containing the "Baidu Apollo Go" tag from January–July 2024. Each post carries metadata: username, IP location (province/municipality granularity), timestamp, repost/comment/attitude counts, and the content text itself. This is the raw input to the entire pipeline.

2. Data Cleaning — A rule-based filter removes two categories of irrelevant content: (a) posts from official media and self-media accounts (identified by matching usernames against a curated list), which lack clear attitudinal stance despite containing the Apollo Go tag; (b) posts manually identified as unrelated in content despite carrying the tag. This stage produces the "initially cleaned" dataset that flows into model inference.

3. Manual Annotation — From the cleaned data, 2,797 posts are randomly sampled for human labeling using the open-source tool doccano. Nine annotators classify each post into one of four categories: "Positive" (excitement, support, anticipation toward Apollo Go), "Negative" (dissatisfaction, disapproval, concern about the service), "Neutral" (objective facts, news reporting, financial information with no emotional bias), or "Drop" (content explicitly irrelevant or so garbled as to be unusable). This labeled subset becomes the training resource for fine-tuning the model.

4. Model Fine-Tuning — The bert-base-chinese pretrained model is fine-tuned on the labeled data. The model takes tokenized Weibo post text as input and produces a three-way classification (positive, negative, neutral) as output. The "Drop" category is used during training to teach the model to recognize irrelevant content, but the final classification stage maps posts to only the three sentiment categories. Hyperparameters are standard: learning rate 1e-5, batch size 8 (both training and evaluation), 5 epochs, Adam optimizer with betas (0.9, 0.999) and epsilon 1e-8, linear learning rate scheduler, seed 42 for reproducibility. The model achieves accuracy of 0.59 on the held-out validation set.

5. Sentiment Classification and Analysis — The fine-tuned model is applied to the full cleaned dataset to produce sentiment labels for every post. These labels — joined with temporal (timestamp) and spatial (IP location) metadata — enable the downstream analyses: temporal trend analysis (tracking sentiment proportions month-by-month and day-by-day in July), spatial distribution analysis (mapping sentiment counts and percentages across Chinese provinces), and thematic analysis (word clouds and keyword examination to characterize what positive and negative commenters discuss).

3.3 Roadmap for the Deep Dive

• First, the data collection mechanism — what the Weibo API provides, what metadata is available, and the structure of individual posts — because understanding the raw material's characteristics is essential for understanding every subsequent processing decision.
• Second, the data cleaning procedure in detail — what is removed, why, and through what mechanisms — since cleaning determines what content the model sees at inference time and shapes the distribution of sentiment the pipeline ultimately measures.
• Third, the annotation protocol — how posts are sampled for labeling, what the labeling categories mean operationally, and how inter-annotator reliability is handled — because the quality of the fine-tuned model depends entirely on the quality of these labels.
• Fourth, the model architecture, fine-tuning procedure, and hyperparameter configuration — what BERT is and why bert-base-chinese was selected, how the labeled data gets split, what accuracy means in this context, and what the 0.59 score implies for downstream results.
• Fifth, the sentiment classification stage — how the trained model maps input text to output labels, and how those labels couple with temporal and spatial metadata to enable the paper's analyses.
• Sixth, the analysis methodology — not new NLP but the framework for interpreting the labeled outputs: temporal grouping (monthly, daily), spatial aggregation (by province), and qualitative keyword extraction (word clouds).

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily an applied empirical study whose technical contribution is not a novel algorithm but rather the construction and application of a domain-specific sentiment analysis pipeline to a novel dataset capturing a live technological deployment. The core idea is that fine-tuning an existing pretrained language model on a modest amount of manually labeled, domain-specific data can produce meaningful sentiment classifications that, when joined with rich metadata (timestamps, geographical coordinates), enable both temporal and spatial analysis of public attitudes toward an operational robotaxi service.

Data Collection: The Weibo API and Post Structure

The dataset originates from the Weibo API service, queried for all posts and comments containing the tag "Baidu Apollo Go" from January through July 2024. The API returns 36,096 records, each carrying the following fields explicated in Table 1: User ID (a unique numeric identifier for the posting account), Content (the full text and/or emoji content of the post), Reposts count (number of times the post was shared), Attitudes count (a Weibo-specific engagement metric roughly equivalent to likes/reactions), Comments count (number of replies), IP location (the geographical location tied to the posting IP address, resolved to the province or municipality level for privacy reasons), and Time (timestamp with minute-level precision, formatted as YYYY/MM/DD HH:MM).

Several properties of this data structure are critical for understanding subsequent methodological choices. First, the IP location is accurate only to province/municipality — meaning Beijing, Shanghai, Chongqing, and Tianjin (all direct-controlled municipalities) appear alongside full provinces (Guangdong, Hubei, etc.) in the spatial analysis. This is a deliberate privacy protection mechanism built into Weibo's data infrastructure, not a choice by the authors. It means that within-province spatial variation (e.g., urban vs. rural areas within Guangdong) cannot be analyzed, but cross-province patterns are robust.

Second, the content field contains the raw Weibo post text, which the paper explicitly identifies as presenting significant NLP challenges: "metaphors, harsh data, and the coexistence of many possible meanings for individual words or phrases" (Section 1). This means that off-the-shelf sentiment models — even Chinese-language ones — may perform poorly without domain-specific adaptation, because colloquial expressions about autonomous vehicles (e.g., Chinese internet slang for Apollo Go, which is colloquially referred to as "萝卜快跑" or "luobokuaipao" — literally "radish run fast") may not appear in the general-domain pretraining corpus.

Third, the temporal distribution is extremely skewed. Figure 1 and Section 4.1 report that 89.56% of all posts (19,597 out of 36,096) are concentrated in July alone, with the remainder scattered across January through June. This is not a methodological limitation but rather the key phenomenon the study analyzes: the July surge represents what the paper calls Apollo Go "becoming a trending topic" on Chinese social media. The implication for methodology is that the model must be trained on posts that span both the pre-surge period (when discussion was quieter and more localized, dominated by Hubei-based posts from areas with operational Apollo Go service) and the post-surge period (when discussion exploded nationally and the content distribution likely shifted toward more diverse and emotionally charged content).

Table 1 sample inspection. Looking at the sample posts in Table 1, we can observe the range of content the pipeline must handle. Post 1 (from Hubei, January 26) is a factual statement about Apollo Go serving an airport route during the Spring Festival travel season — likely neutral. Post 2 (Hubei, February 27) reports the milestone of autonomous taxis crossing the Yangtze River — again factual. Post 3 (Beijing, March 23) expresses personal enthusiasm: "Immersive experience with autonomous vehicles, super cool!" — clearly positive. Post 5 (Anhui, July 11) expresses a more complex sentiment: support for Apollo Go framed as a rejection of traditional taxi driver behavior ("No more dealing with taxi drivers rolling their eyes and complaining about short trips") — positive toward Apollo Go, negative toward the incumbent service. This mixture of factual reporting, personal experience sharing, emotional reactions, and comparative commentary within a single dataset is precisely why fine-tuning on domain-specific labeled data is necessary rather than relying on a general-purpose sentiment classifier.

Data Cleaning: Removing Noise Without Losing Signal

The original dataset contains two distinct categories of content that the paper identifies as needing removal before sentiment analysis can proceed meaningfully. These are conceptually different problems requiring different cleaning strategies.

Category 1: Official media and self-media accounts. These are posts where the content carries the "Baidu Apollo Go" tag (hence their inclusion in the API retrieval) but the posting account belongs to an official news organization, a corporate account, or a self-media influencer whose content describes facts rather than expressing personal attitudes. The paper states that these records are "dropped directly due to their lack of clear attitude as their content focused on describing facts" (Section 2.1). The cleaning mechanism is rule-based username matching: the authors maintain a list of known official media accounts and match each post's User ID against this list.

This is an important methodological choice. One might argue that these posts do express attitudes — just implicitly, through framing choices, source selection, or story emphasis. However, the paper's goal is to measure public sentiment (ordinary users' reactions), not institutional communication. Including official media posts would systematically bias the sentiment distribution toward neutral, since institutional reporting typically presents itself as objective, and would obscure the genuine attitudinal signal from everyday users that the study seeks to capture.

Category 2: Tag-matched but content-irrelevant posts. Some users tag their posts with trending topics (including "Baidu Apollo Go") to increase visibility even though the post content is unrelated to Apollo Go. Additionally, some users post long-form content (blogs, daily diaries) in which Apollo Go is mentioned in passing but is not the subject of the post. These posts pose a harder cleaning challenge because they cannot be identified by username alone — the account is an ordinary user, but the specific post's relationship to Apollo Go is tangential.

The paper takes a hybrid approach to this problem. For long texts where Apollo Go appears but is not the focus, 20% are manually annotated to determine whether the post contains a genuine attitudinal expression about the service (in which case it is retained) or whether it merely mentions Apollo Go in a list of daily activities (in which case it is dropped). The paper states: "Once the long texts are manually annotated, feature extraction methods can be applied to the remaining records" — though the specific feature extraction method is not detailed, the implication is that the manually annotated long texts serve as training data for an automated classifier that can then process the remaining 80%.

The "Drop" category. Posts that survive the username-based media filter but are identified as irrelevant through manual inspection are labeled as "Drop" during the annotation phase. The labeled dataset includes 320 "Drop" samples out of 2,697 total (Table 2), representing approximately 11.9% of the annotated set. This category is used during model fine-tuning to teach the classifier to distinguish irrelevant content from attitudinal content, preventing the model from assigning random sentiment labels to off-topic posts during inference on the full dataset.

What remains. After cleaning, the remaining posts are those where an ordinary user explicitly discusses Apollo Go in a way that expresses or implies an attitudinal stance (positive, negative, or neutral description of facts relevant to the service). The paper does not report the exact number of posts that survive cleaning — the 36,096 figure is pre-cleaning — but the downstream analysis operates on the cleaned subset.

Manual Annotation: Building the Training Signal

The paper's model is not trained on an existing sentiment dataset; it is trained on data specifically labeled for this study's domain and categories. This is critical because general Chinese sentiment lexicons or pretrained sentiment classifiers are calibrated on domains like product reviews, movie comments, or news articles — distributions that differ substantially from social media discourse about autonomous ride-hailing.

Annotation tool and personnel. Nine individuals participated in the annotation process using doccano, an open-source text annotation tool. The paper does not report inter-annotator agreement metrics (e.g., Cohen's kappa, Fleiss' kappa, or percentage agreement), which is a notable omission — the reliability of the training labels cannot be assessed from the reported information alone.

Sampling and final labeled set. From the cleaned data, 2,797 posts were selected for manual annotation. After removing records that annotators identified as "excessively long and meaningless texts," 2,697 posts remained for model training and evaluation. Table 2 reports the distribution across the four categories:

The dataset is moderately imbalanced — positive posts outnumber negative posts by approximately 1.73:1. This class imbalance is not corrected through resampling or weighted loss during training (no such procedures are mentioned), meaning the fine-tuned model may have a slight bias toward predicting positive sentiment, which could modestly inflate the positive proportion in the final classification results.

Operational definitions of the categories. The paper provides explicit classification standards that the annotators used:

• Positive: "Texts exhibiting emotions such as excitement, enthusiasm, support, and encouragement... indicating support and anticipation for Baidu Apollo Go and similar technological services, with the belief that further deep testing and early commercial operation are warranted" (Section 2.2). The key criterion is that the text expresses a favorable stance toward Apollo Go or autonomous driving technology and implies that continued development/deployment is desirable.

• Negative: "Texts expressing dissatisfaction, disgust, disapproval, or abusive language... indicating a negative attitude towards Baidu Apollo Go and similar services, suggesting that the development of this technology should not continue at this time" (Section 2.2). The key criterion is opposition or concern sufficient to imply that deployment should slow or stop.

• Neutral: "Other content, such as news articles, media reports, and financial information... they merely stated objective facts without emotional bias" (Section 2.2). The key criterion is that the text conveys information without an evaluative stance — it does not advocate for or against Apollo Go.

• Drop: Content explicitly identified as irrelevant to Apollo Go or so garbled as to be unsuitable for sentiment classification.

Train-validation split. The labeled dataset of 2,697 posts is split into training and validation sets in a 4:1 ratio, yielding approximately 2,158 training samples and 539 validation samples. This is a relatively small training set by modern deep learning standards — fine-tuning a 110-million-parameter BERT model on ~2,000 examples risks overfitting, which may partly explain the reported accuracy of 0.59 (substantially lower than what bert-base-chinese typically achieves on well-resourced sentiment tasks).

Why manual annotation over pre-existing sentiment resources? The paper does not explicitly argue this, but the choice is strategically sound for the domain. Chinese social media discourse about autonomous vehicles contains domain-specific expressions (e.g., colloquial names for Apollo Go, references to specific incidents like the Apollo Day 2024 event, comparisons with traditional taxi driver behaviors) that general sentiment lexicons would not capture. For instance, a post reading "Fully support Baidu Apollo Go! No more dealing with taxi drivers rolling their eyes and complaining about short trips" (Table 1, User ID 5953250708) expresses a complex sentiment that combines positive affect toward Apollo Go with negative affect toward traditional taxis — a nuance that a dictionary-based sentiment scorer would likely misclassify. Training on domain-specific labels ensures the model learns these context-dependent sentiment signals.

Model Architecture and Fine-Tuning: The bert-base-chinese Pipeline

What BERT is. The model used is BERT (Bidirectional Encoder Representations from Transformers), specifically the bert-base-chinese variant pretrained on Chinese text and available through the Hugging Face model hub. BERT is a transformer-based neural network that processes text bidirectionally — when encoding a token, it attends to both the tokens before and after it in the sequence, unlike autoregressive models (like GPT) that only attend to previous tokens. This bidirectional context is critical for sentiment analysis because the sentiment of a word often depends on words that follow it (consider "not good" vs. "good" — a left-to-right model seeing "not" cannot yet know it negates "good," but BERT can attend to both simultaneously).

Architecture specifications (implicit). While the paper does not enumerate BERT's architecture in detail, bert-base-chinese uses the standard BERT-base configuration: 12 transformer layers, 768 hidden dimensions, 12 attention heads, approximately 110 million parameters, and a vocabulary of 21,128 Chinese characters/tokens. The model was pretrained on Chinese Wikipedia and other Chinese corpora using the masked language modeling (MLM) objective: during pretraining, 15% of input tokens are randomly masked, and the model learns to predict the original tokens from context.

Why bert-base-chinese specifically. The paper states three justification criteria (Section 3): "its proven effectiveness in handling Chinese text, its ability to capture contextual nuances, and its strong performance in various natural language processing tasks." The bidirectional attention mechanism is particularly relevant for Chinese, where word boundaries are not marked by spaces and the meaning of a character sequence depends on contextual disambiguation. For example, in simplified Chinese social media, the characters for "Apollo Go" (百度萝卜快跑) could appear in various abbreviated or colloquial forms whose sentiment implications depend on surrounding context — something BERT's token-level bidirectional encoding handles naturally.

Fine-tuning procedure. Fine-tuning is the process of taking the pretrained BERT model and continuing training on the domain-specific labeled data. The pretrained model already understands Chinese syntax and general semantics; fine-tuning adapts these representations to the specific task of classifying Weibo posts about Apollo Go into sentiment categories. This is a form of transfer learning that the paper justifies by noting that "fine-tuning techniques preserve the pretrained knowledge of these large models while significantly reducing the amount of labeled data and training time required" (Section 1, citing Christofidellis, 2023).

The fine-tuning adds a classification head — a small feedforward neural network — on top of BERT's final hidden state corresponding to the [CLS] token (a special token prepended to every input sequence, whose output representation is designed to capture sequence-level information). This classification head maps the 768-dimensional [CLS] embedding to either 3 or 4 logits (depending on whether "Drop" is included in the classification task; the paper is somewhat ambiguous on this point but the downstream analysis uses only positive/negative/neutral, suggesting that the model either classifies into 4 categories with Drop posts removed post hoc, or classifies into 3 categories with Drop posts excluded during training).

Hyperparameter configuration (verbatim from Section 3). The paper reports the following training hyperparameters:

Learning rate: 1e-05
Training batch size: 8
Evaluation batch size: 8
Seed value: 42
Optimizer: Adam with betas (0.9, 0.999) and epsilon 1e-08
Learning rate scheduler: linear
Number of epochs: 5

These are standard, conservative fine-tuning settings. The learning rate of 1e-5 is an order of magnitude smaller than typical pretraining learning rates — this ensures that the updates during fine-tuning are subtle, "making subtle adjustments to the pre-trained BERT model's parameters" and adapting to the specific task "without significantly disrupting the representations it has already learned" (Section 3). The batch size of 8 is small, appropriate for the modest training set size (~2,158 samples), and the linear learning rate scheduler gradually reduces the learning rate over the 5 epochs, which helps the model converge smoothly.

The loss function (not stated but necessarily present). While the paper does not explicitly provide the training loss, the standard objective for multi-class classification with BERT is cross-entropy loss over the softmax-normalized logits. For a single training example with ground-truth class $y$ (represented as a one-hot vector over the sentiment categories) and model-predicted probabilities $\hat{y}$ (from softmax over the logits), the loss is:

$\mathcal{L} = -\sum_{c=1}^{C} y_c \log(\hat{y}_c)$

where $C$ is the number of classes (3 or 4, depending on whether "Drop" is included in the classification head), $y_c \in \{0, 1\}$ is a binary indicator for whether class $c$ is the correct label, and $\hat{y}_c \in [0, 1]$ is the model's predicted probability for class $c$, satisfying $\sum_c \hat{y}_c = 1$.

What this loss computes in operational terms: For each training post, the model produces a probability distribution over the sentiment categories. The loss penalizes the model proportionally to the negative log of the probability it assigns to the correct category — if the model assigns probability 0.9 to the correct label, the loss is $- \log(0.9) \approx 0.105$; if it assigns probability 0.1, the loss is $- \log(0.1) \approx 2.303$. The Adam optimizer uses the gradient of this loss with respect to all 110 million parameters to update the model weights in the direction that would increase the probability assigned to correct labels on future similar inputs.

Why cross-entropy: This is the standard maximum-likelihood objective for classification — minimizing cross-entropy is equivalent to maximizing the probability the model assigns to the correct label under a categorical distribution. It is the correct objective when the goal is calibrated probability estimates (rather than, say, just correct top-1 predictions), because it heavily penalizes confident mistakes — a model that assigns 0.99 probability to the wrong class incurs a much larger loss than one that assigns 0.6, encouraging conservative probability estimates.

Model accuracy. The paper reports that after fine-tuning, "the model achieved an accuracy of 0.59" (Section 3). Accuracy here is the fraction of validation-set posts for which the model's highest-probability class matches the human-assigned label. An accuracy of 0.59 means that on the held-out validation set of ~539 posts, the model correctly classifies approximately 318 posts and misclassifies approximately 221. This is substantially better than random guessing (which would achieve ~0.33 for three balanced classes or ~0.25 for four classes), but is moderate by the standards of modern sentiment analysis — BERT fine-tuned on large, high-quality sentiment datasets routinely achieves accuracy above 0.85–0.90.

What explains the 0.59 accuracy? Several factors likely contribute. First, the training set is small (~2,158 examples for a 110M-parameter model), creating genuine risk of overfitting and limiting the model's ability to learn robust sentiment features. Second, the annotation task is inherently ambiguous — the boundary between "neutral" and "mildly positive" or "neutral" and "mildly negative" is subjective, meaning that even perfect inter-annotator agreement would be unlikely and the model's ceiling accuracy is below 1.0. Third, Chinese social media text presents genuine linguistic challenges (informal language, creative character usage, implicit sentiment expressed through emoji or cultural references) that even a pretrained model may struggle to fully resolve from limited domain-specific examples.

Implications of 0.59 accuracy for downstream analysis. The paper acknowledges this accuracy figure but does not deeply interrogate its implications for the reliability of the reported findings. When the model classifies 36,096 posts with 59% accuracy, the resulting sentiment distribution contains non-trivial noise — roughly 4 out of every 10 classifications are expected to be wrong. However, if the errors are approximately symmetric (i.e., the model confuses positive and negative at roughly equal rates) and the true sentiment distribution is not too extreme, the aggregate proportions may still be reasonably reliable — a phenomenon known as "classification with symmetric error." The paper does not perform error analysis to characterize whether misclassifications are symmetric, which leaves uncertainty about whether the reported temporal and spatial sentiment patterns reflect genuine attitudinal variation or are artifacts of systematic model bias (e.g., the model might systematically misclassify neutral posts with technical vocabulary as positive, inflating positive proportions in provinces that discuss Apollo Go's technology in detail).

Sentiment Classification: Applying the Model at Scale

Once fine-tuned, the model is applied to the full cleaned dataset (the ~36,096 posts minus those removed during cleaning). For each post, the pipeline is:

1. Tokenization: The post's Chinese text is segmented into tokens using BERT's pretrained WordPiece tokenizer for Chinese. The tokenizer handles Chinese characters individually (since Chinese lacks explicit word boundaries), prepends the [CLS] token, and appends the [SEP] token (used to mark the end of the sequence). Posts longer than BERT's maximum sequence length of 512 tokens are truncated; posts shorter than this are padded with [PAD] tokens to a uniform length within the batch.

2. Forward pass: The tokenized input is fed through BERT's 12 transformer layers. At each layer, multi-head self-attention computes contextual representations by allowing each token to attend to every other token in the sequence. The output at the final layer for the [CLS] token — a 768-dimensional vector — serves as the sequence representation.

3. Classification: The classification head (a linear layer mapping 768 dimensions to the number of classes, followed by softmax) converts the [CLS] representation into a probability distribution over sentiment categories.

4. Label assignment: The category with the highest predicted probability is assigned as the post's sentiment label. Importantly, only three categories — positive, negative, and neutral — are used in the final analysis; posts the model might classify as "Drop" are presumably either excluded or the model's classification head is configured for only three categories (the paper is ambiguous on this point).

Output structure. The classification stage produces a dataset where each of the cleaned posts now carries, in addition to its original metadata (user ID, timestamp, IP location, engagement counts), a machine-assigned sentiment label (positive, neutral, or negative). This structured dataset is the input to all downstream analyses.

Temporal and Spatial Analysis Framework

The final technical component is the analytical framework that transforms labeled posts into the temporal and spatial patterns presented in Section 4. This is not machine learning but rather a series of aggregation and visualization procedures whose methodological choices affect what conclusions can be drawn.

Temporal aggregation. Posts are grouped by time unit at two granularities: monthly (for the January–July overview) and daily (for the July zoom-in analysis). For each time unit, the number of posts in each sentiment category is counted, and proportions are computed. The paper presents these as both raw counts (Tables 4 and 5) and stacked area charts (Figures 5a, 5b). Two analytical decisions are noteworthy:

• The January–June period is treated as a single analytical phase and contrasted against July. This is driven by the data: July contains 89.56% of all posts, making the pre-July period a natural "pre-trending" baseline. However, this also means that the pre-July sentiment estimates are based on much smaller samples (only 1,771 posts across six months), introducing higher variance in those estimates relative to the July estimates (19,597 posts in one month).
• The daily analysis in July reveals a surge structure: posts begin low (7 on July 1), rise through the first week (~69 by July 6), then explode around July 8–12 (970 on July 8, 2,826 on July 10, 6,226 on July 12) before declining. The paper uses this structure to analyze how sentiment composition changes as the volume of discourse increases — finding that negative sentiment proportion grows during and after the peak.

Spatial aggregation. Posts are grouped by IP location (province/municipality level). For each province, three metrics are computed: the count of posts in each sentiment category, and the percentage distribution of sentiment categories within that province. These are visualized as:

• Side-by-side bar charts and choropleth maps showing raw counts (Figure 9 for the full period, Figure 11 for daily snapshots in July)
• Choropleth maps showing percentage distributions (Figure 10 for the full period, Figure 12 for daily snapshots in July)

The choropleth visualization choice — using a tricolor scheme where blue represents negative, green neutral, and red positive — enables rapid visual identification of provinces where sentiment deviates from the national average. The paper explicitly notes that "the eastern coastal provinces show a higher level of discussion intensity" (Section 4.2), and the spatial analysis establishes the correlation between operational deployment regions and discussion volume.

Word cloud generation. For qualitative analysis of what positive and negative commenters discuss, the paper generates word clouds from the content of posts in each sentiment category (Figure 13). The word cloud is a frequency-based visualization where font size is proportional to word occurrence count. This is not a deep NLP technique — it is essentially a unigram frequency counter with stop-word removal — but it serves a useful purpose: quickly identifying the dominant themes in each sentiment class without requiring manual reading of thousands of posts. The paper reports that positive comments center on words like "driverless," "automatic," "technique," "technology," and "development," while negative comments center on "response," "unemployment," "market," and "problem." The word cloud methodology does not account for n-gram phrases, negation scope, or word sense disambiguation — "response" (响应) could refer to user satisfaction with Apollo Go's response to their needs or to government regulatory response, and the unigram approach cannot distinguish these.

Why these analyses rather than more sophisticated NLP? The paper's analytical framework is deliberately simple: count, group, visualize. This reflects the paper's positioning as a transportation policy study using NLP as an instrument rather than an NLP contribution per se. More sophisticated analyses — topic modeling (LDA), aspect-based sentiment analysis, named entity recognition for specific concerns, or geospatial regression to identify drivers of provincial sentiment variation — could extract richer signals from the data, but the paper's goal is to establish the temporal and spatial structure of public sentiment at a high level, which counts and proportions suffice to demonstrate.

Summary of Design Choices and Justifications

• Manual annotation with four categories (including "Drop") rather than relying on existing sentiment lexicons or zero-shot classifiers: necessary because domain-specific expressions (colloquial names for Apollo Go, context-dependent sentiment about technology vs. labor) would be missed by general-domain tools, and the "Drop" category is needed to handle irrelevant posts that pass the username filter but are not about Apollo Go.

• Fine-tuning bert-base-chinese rather than using a simpler model (e.g., SVM with bag-of-words features, LSTM): BERT's bidirectional attention captures contextual nuances that matter for sentiment (e.g., negation, sarcasm, comparative statements about Apollo Go vs. traditional taxis) that would be invisible to unidirectional or bag-of-words approaches.

• Conservative hyperparameters (low learning rate, small batch size, 5 epochs): appropriate for the small training set size; aggressive training would rapidly overfit the ~2,000 training examples, memorizing them rather than learning generalizable sentiment features.

• Binary sentiment classification (positive/negative/neutral) rather than fine-grained sentiment (1–5 scale or emotion categories): simplifies the annotation task (making it feasible for nine annotators without extensive training) and focuses the analysis on the policy-relevant dimension of public acceptance vs. opposition rather than subtle emotional gradations.

• Rule-based removal of official media rather than training the model to distinguish institutional from personal sentiment: simpler, more reliable, avoids the risk of the model learning to associate certain formal vocabulary with neutral sentiment and misclassifying formal-but-opinionated personal posts.

• Word clouds rather than topic models for qualitative content analysis: the paper prioritizes interpretability and visual communication over analytical depth; word clouds are immediately legible to policymakers and planners who may lack NLP expertise, while an LDA topic model would require explaining latent topic representations and their interpretation.

4. Key Insights and Innovations

Innovation 1: The Intervention-Based Natural Experiment — Using a Viral Trend as a Temporal Breakpoint

The paper's most intellectually distinctive move is not its NLP pipeline but its research design: it treats Apollo Go's July 2024 viral surge on Chinese social media as a natural experiment with a pre-post intervention structure, where the "intervention" is the sudden, massive expansion of public awareness triggered by media coverage and the Apollo Day 2024 event (May 15, 2024, Wuhan). The study does not merely observe sentiment over time; it structures its entire analytical framework around the contrast between the pre-trending period (January–June: 1,771 posts, quiet, localized discussion dominated by provinces where Apollo Go actually operates) and the post-trending period (July: 19,597 posts, national in scope, 89.56% of all data concentrated in a single month).

This is not how most social media sentiment studies in the transportation domain are structured. The dominant approach — exemplified by Jiang and Zheng (2021) studying AV sentiment on Twitter, Ding et al. (2021) analyzing how external events influence AV-related tweets, and Chen and Tomblin (2021) combining Reddit, deliberation transcripts, and surveys — collects data continuously and analyzes temporal trends as smooth variations or event-correlated fluctuations. These studies typically look for sentiment shifts correlated with news events (e.g., an accident, a regulatory announcement) but lack the sharp phase-change structure that the Apollo Go dataset naturally provides: a service that was quietly operating in the background for six months, suddenly becoming a national trending topic and generating an order-of-magnitude increase in discussion volume.

The paper's pre-post framing enables a specific diagnostic claim that is more powerful than a general "sentiment evolved over time" observation. The Abstract states it directly: "From January to July, public sentiment was mostly positive, but negative comments began to rise after it became a hot topic on July 21." This is not a vague trend — it is a phase-specific finding: the sentiment distribution of pre-trending discourse (dominated by early adopters, local users in operational cities, and technology enthusiasts) is qualitatively different from post-trending discourse (dominated by a broader national audience encountering the technology through media rather than personal experience). The paper's temporal analysis (Section 4.1, Figures 5a and 5b, Tables 4 and 5) operationalizes this: the surge from ~742 posts in May and 685 in June to 19,597 in July is not merely a volume increase — it represents a shift in who is talking and consequently what they say.

What makes this a genuine innovation rather than just a property of the dataset. The intellectual contribution is recognizing that the data's temporal structure enables analysis that survey-based research cannot replicate. In a cross-sectional survey, you ask respondents at one point in time about their attitudes; you cannot observe how the composition of discussants changes when a topic goes viral, because the survey sample is fixed. In a continuous social media monitoring study with relatively stable volume, you can track sentiment shifts but cannot isolate the effect of audience expansion. The Apollo Go case — with its step-function volume increase and documented shift from localized operational discourse to national public debate — provides empirical traction on a question that is both theoretically important and methodologically difficult: how does public sentiment toward a transportation technology change when the audience expands from early adopters and local stakeholders to the general national public? The finding that negative sentiment proportion increases as the audience broadens (Section 4.1) has direct implications for how transportation agencies and service providers should interpret early positive sentiment from pilot deployments — it may not generalize.

Evidence anchoring. The claim is supported by the temporal analysis in Figures 5a and 5b, which show the January–June period having roughly comparable proportions of positive vs. negative sentiment across months, while the July daily breakdown (particularly July 10–12, the peak days) shows negative sentiment expanding both in absolute count and in proportion relative to positive. The raw numbers in Table 5 are striking: July 1 started with 4 positive vs. 2 negative posts (a 2:1 ratio), but by July 12 (6,226 total posts), the ratio had compressed to 2,429 positive vs. 2,273 negative — nearly 1:1. This shift in the sentiment ratio concurrent with the volume surge is the empirical basis for the pre-post phase-change claim.

Innovation 2: Geographic Sentiment Heterogeneity as a First-Order Finding — The "Xinjiang-Tibet Divergence"

The paper's second conceptual contribution is establishing that public sentiment toward an identical autonomous ride-hailing service can vary dramatically across subnational regions within a single country, and that this variation is not merely a function of whether the service operates in that region. This is not a minor empirical finding; it challenges an implicit assumption in much of the AV acceptance literature that "public sentiment" within a national market is broadly uniform or varies primarily along individual-level demographic dimensions (age, income, tech-savviness).

Prior work on AV public acceptance has primarily operated in one of two modes. Survey-based studies (Bansal et al., 2016; Montoro et al., 2019; Roche-Cerasi, 2019) sample individuals and identify demographic and psychographic predictors of acceptance — being male, younger, urban, high-income, and tech-savvy all correlate with AV enthusiasm. These studies capture individual-level variation but typically treat geography as a control variable (urban vs. rural) rather than as a first-order axis of analysis. Social media studies in Western contexts (Ding et al., 2021; Jiang and Zheng, 2021) can sometimes infer user location from profile text or tweet content, but as the paper explicitly notes, platforms like Twitter do not provide verified IP-level location data at the granularity that Weibo does — meaning location inference is noisy and coarse.

The Weibo IP location feature, accurate to province/municipality level, enables the paper to produce something genuinely novel: a province-level sentiment map of China for Apollo Go discourse (Figures 9 and 10). And the map reveals a pattern that is not simply "provinces with Apollo Go are more positive":

• The eastern coastal provinces that dominate discussion volume (Guangdong, Beijing, Shandong, Zhejiang, Jiangsu) show mixed sentiment — high absolute counts of both positive and negative posts, reflecting the national debate playing out in population centers.
• Xinjiang and Qinghai — western provinces where Apollo Go does not operate at all — are "predominantly shaded in deep red" (Section 4.2), indicating strong positive sentiment. The paper characterizes these provinces as expressing "optimism about the unstoppable trend of autonomous driving and anticipation for its development."
• Tibet and Gansu — also western, non-operational provinces — are "primarily shaded in deep blue," reflecting "concerns about the impact of autonomous vehicles on traditional taxi and ride-hailing services."

The Xinjiang-Tibet divergence is the paper's most striking specific finding. These are neighboring regions (both in western China, both without Apollo Go operations, both likely exposed to the same national media coverage), yet they exhibit diametrically opposite sentiment distributions. This cannot be explained by differential exposure to the service (neither has it) or by individual demographic factors alone (unless there is an implausibly sharp demographic discontinuity at the provincial border). It implies that province-level cultural, economic, or informational factors are shaping aggregate sentiment in ways that individual-level models would not predict.

What makes this a conceptual advance rather than just a map. The paper does not merely produce a pretty choropleth; it uses the geographic pattern to make an argument about the mechanisms of public sentiment formation. The finding that discussion intensity correlates with operational provinces (Table 6, Figure 8) is intuitive — people talk more about what affects them directly. But the finding that non-operational provinces show sharp sentiment polarization is counterintuitive and theoretically significant: it suggests that in regions without direct experience with the technology, sentiment is shaped by second-order factors — media framing, economic anxiety about industries that might be disrupted, cultural attitudes toward technological change, or local labor market structures. The paper does not identify which of these mechanisms drives the Xinjiang-Tibet divergence (it lacks the demographic and economic data to do so), but establishing that such divergence exists at this granularity is a contribution in itself — it changes the question from "what predicts individual AV acceptance?" to "what predicts regional AV sentiment patterns, and how do those patterns interact with deployment geography?"

Evidence anchoring. The spatial maps in Figures 9 (raw counts) and 10 (percentages) are the primary evidence. Figure 10, in particular, uses a tricolor scheme where the proportion of red (positive), green (neutral), and blue (negative) is visible for each province at a glance — and the visual contrast between deep-red Xinjiang/Qinghai and deep-blue Tibet/Gansu in the western region is stark. The daily snapshots in Figure 12 (July 8–14) provide temporal granularity, showing that this geographic heterogeneity persists throughout the viral surge period and is not an artifact of aggregation across time.

Innovation 3: Domain-Specific Sentiment Category Structure — Separating "Technology Enthusiasm" from "Labor Anxiety" as Distinct Attitude Clusters

The paper's third contribution operates at the intersection of methodology and substantive finding: by combining quantitative sentiment classification with qualitative word cloud analysis (Figure 13), it demonstrates that positive and negative sentiment toward Apollo Go are not merely opposite ends of a single attitudinal spectrum but are driven by fundamentally different thematic concerns. Positive sentiment clusters around technology experience (keywords: "driverless," "automatic," "technique," "technology," "development"), while negative sentiment clusters around socioeconomic threat (keywords: "unemployment," "market," "problem") and operational concerns (keyword: "response").

This is a conceptual refinement over standard sentiment analysis, which typically treats "positive" and "negative" as symmetric opposites on a single valence dimension. The word cloud decomposition reveals that the two sentiment poles are anchored in qualitatively different domains: positive sentiment is about what the technology enables (novel experiences, convenience, a glimpse of the future), while negative sentiment is about what the technology threatens (jobs, the existing economic order, safety). This is not merely "some people like it and some don't" — it is that the basis of liking and disliking are orthogonal. A person could simultaneously find the technology exciting (positive valence on the technology-experience dimension) and worry about its labor impacts (negative valence on the socioeconomic-threat dimension) — and the sentiment classification forces a single label, potentially obscuring this ambivalence.

Prior survey research (Tussyadiah et al., 2017; Bansal et al., 2016; Nikitas et al., 2021) has documented that AV attitudes are multidimensional — trust in technology, perceived safety, concern about job displacement, and excitement about convenience all factor into overall acceptance. But these studies typically use structured questionnaires that pre-specify the dimensions of concern and ask respondents to rate each one. The paper's word cloud approach is complementary: it allows the themes to emerge from the data itself, revealing which concerns organically dominate in spontaneous public discourse. The fact that "unemployment" dominates the negative word cloud while "safety" is present but not as prominent (the paper notes safety concerns but the word cloud emphasizes labor displacement) suggests that, in the Chinese social media context, labor market anxiety is the primary driver of opposition — a finding that differs from Western contexts, where the paper notes that privacy and data security concerns are more prominent (Section 5, citing Chen and Tomblin, 2021).

The significance beyond this specific paper. This thematic decomposition provides actionable guidance for policymakers and service providers that a simple "X% positive, Y% negative" breakdown would not. If negative sentiment is primarily driven by labor displacement concerns rather than safety fears, the appropriate response is not safety demonstrations (which address the wrong concern) but rather workforce transition programs, public communication about job creation in adjacent sectors, or policy frameworks for managing the labor market transition. The paper gestures toward this in its discussion (Section 5): "Effective strategies might include enhancing public outreach, conducting trial rides, and refining legal frameworks to mitigate potential negative impacts, such as job losses or liability issues in accidents" — the word cloud analysis tells us that "job losses" should be the priority, not just one item on a list.

Evidence anchoring. Figure 13 provides the direct evidence: two side-by-side word clouds where the visual salience of "unemployment" (失业) in the negative cloud and "technology" (技术) / "driverless" (无人驾驶) in the positive cloud makes the thematic asymmetry immediately apparent. The paper supplements this with qualitative inspection of posts: "most of the positive comments focused on the discussion of technology applications... personal experience of Apollo Go, as well as the positive news and reports," while "negative discussions focused mainly on the concerns about job displacement... and the safety issues" (Section 4.3). The asymmetry is not perfectly clean — safety appears in both clouds and some positive posts mention practical benefits — but the dominant themes are clearly divergent.

Innovation 4: The "Inverse Local Discussion" Pattern — How Viral Attention Decouples Discourse Geography from Deployment Geography

A fourth, more subtle contribution emerges from the spatial dynamics of the July surge. In the January–June pre-trending period, the discussion was dominated by Hubei province (383 posts, ranked 1st) and Guangdong (362 posts, ranked 2nd) — provinces where Apollo Go operates commercially and where users had direct experience with the service (Table 6, Figure 8a). This is exactly what one would expect: discussion of a localized service is concentrated where the service exists.

But in July, the geography of discussion shifts dramatically in a way that is not merely a scaled-up version of the pre-existing pattern. Guangdong jumps to 1st place (3,615 posts), Beijing to 2nd (2,483 posts), and Hubei drops to 3rd (1,477 posts). Moreover, the paper notes that "the level of discussion abroad has also surpassed that in Wuhan" (Section 4.2). Provinces with no Apollo Go operations — Zhejiang, Jiangsu, Shandong, Henan, Sichuan, Fujian — all appear in the July top 10 with post counts in the hundreds, collectively generating more discussion than operational provinces.

This is what might be termed an inverse local discussion pattern: when a localized service goes viral nationally, the geography of online discourse shifts away from the original deployment centers toward population centers and media hubs, even if those hubs lack direct experience with the service. Hubei's absolute post count increases (383 → 1,477), but its relative share collapses as millions of users in Guangdong, Beijing, and international locations — who cannot hail an Apollo Go and have never ridden in one — join the conversation.

Why this is an innovation rather than an obvious observation. The pattern reveals a structural feature of how new mobility technologies enter public consciousness that is not captured by standard diffusion-of-innovation models. In the classic Rogers diffusion framework, awareness spreads from early adopters (who have direct experience) to the broader population through communication channels — but the geography of awareness is assumed to track adoption geography, with awareness highest where adoption is occurring. The Apollo Go Weibo data suggests something different: when a technology's social media virality outpaces its geographic availability, a majority of the online discourse is generated by people who cannot use the service. The sentiment of these non-users — shaped entirely by media coverage, word-of-mouth, and ideological priors — becomes the dominant public signal, potentially drowning out the more nuanced, experience-based sentiment of actual users in Hubei.

This has important practical implications. A service provider monitoring social media sentiment to gauge public acceptance might observe the July surge and conclude that national sentiment is increasingly mixed (which it is in the aggregate). But that aggregate conceals a crucial compositional effect: the post-July discourse includes a large fraction of discussants whose sentiment is not grounded in experience. If these non-user sentiments are systematically different from user sentiments (and the paper's data suggests they may be — the pre-July Hubei-dominated discourse was proportionally more positive), then "national sentiment" as measured by social media is a noisy and potentially misleading signal for actual user acceptance.

Evidence anchoring. Table 6 provides the raw ranking and count data that establishes the spatial pattern shift. Figure 8 presents the same information as side-by-side choropleth maps (January–June vs. July), making the geographic decoupling visually apparent: the January–June map shows a strong Hubei-Guangdong-Beijing concentration (all operational provinces), while the July map shows a much more spatially diffuse pattern with high intensity along the entire eastern seaboard. The paper's commentary (Section 4.2) explicitly notes the interpretive implication: "This might be caused by the discussion in Hubei becoming localized and primarily happening offline" — suggesting that as national discourse exploded, the original local user conversation became a smaller component of the measured signal, even as (presumably) actual usage in Hubei continued or expanded.

Limitation of this claim. The paper does not have data on actual Apollo Go ridership or user demographics, so the decoupling claim is inferential: it rests on the assumption that Weibo posters outside operational provinces are unlikely to have direct Apollo Go experience. This is reasonable but not proven — some Guangdong-based Weibo users might have traveled to Wuhan and used Apollo Go, and international discussion could include Chinese diaspora who experienced the service during visits. The paper acknowledges this limitation implicitly through its focus on IP location rather than verified usage.

5. Experimental Analysis

Evaluation Methodology

• Dataset. The dataset consists of 36,096 Weibo posts and comments containing the "Baidu Apollo Go" tag, collected via the Weibo API from January through July 2024. Each post carries metadata: a unique User ID, the full Content text (including emoji), IP location resolved to province/municipality level (for privacy protection), Time with minute-level precision (YYYY/MM/DD HH:MM format), and engagement metrics (Reposts count, Attitudes count, Comments count). From this full dataset, 2,797 posts were sampled for manual annotation using the doccano tool. After removing "excessively long and meaningless texts," 2,697 annotated posts remained and were split into training and validation sets in a 4:1 ratio (approximately 2,158 training, 539 validation). The distribution across annotation categories is reported in Table 2: Positive (1,089 samples, 40.4%), Neutral (659, 24.4%), Negative (629, 23.3%), and Drop (320, 11.9%). The model is then applied to the full cleaned dataset for inference; the paper does not report the exact post-cleaning dataset size, but it is the 36,096 minus those removed by username-based filtering of official media accounts and manual removal of long irrelevant posts.

• Base model(s). The model is bert-base-chinese, a pretrained BERT model available on Hugging Face specifically designed for Chinese text. It uses the standard BERT-base architecture: 12 transformer layers, 768 hidden dimensions, 12 attention heads, approximately 110 million parameters, and a vocabulary of roughly 21,128 Chinese characters/tokens. The model was selected for "its proven effectiveness in handling Chinese text, its ability to capture contextual nuances, and its strong performance in various natural language processing tasks" (Section 3). The choice is justified by the specific challenges of Chinese social media text — metaphors, ambiguous word boundaries, domain-specific neologisms like colloquial names for Apollo Go (e.g., "萝卜快跑" / "luobokuaipao") — which require bidirectional context for accurate sentiment disambiguation.

• Metrics. The primary evaluation metric is classification accuracy on the held-out validation set, computed as the fraction of posts for which the model's highest-probability predicted sentiment class matches the human-assigned label. The paper reports a single accuracy figure of 0.59 after fine-tuning (Section 3). This accuracy is computed over the three relevant sentiment categories (positive, neutral, negative) plus the "Drop" category during training, though the downstream analysis uses only positive/negative/neutral. The paper does not report per-class precision, recall, or F1 scores, does not provide a confusion matrix, and does not report baseline accuracy from a majority-class or random classifier — all of which would help contextualize the 0.59 figure. In the downstream temporal and spatial analyses, the primary reported statistics are raw post counts per sentiment category per time period or province, and sentiment proportions (percentages). No confidence intervals or statistical significance tests are reported for any temporal or spatial comparisons.

• Baselines. The paper does not compare its fine-tuned BERT model against any alternative sentiment analysis approach. There is no baseline comparison with: (a) a lexicon-based sentiment scorer (e.g., a Chinese sentiment dictionary applied to the same posts); (b) an off-the-shelf pretrained sentiment classifier for Chinese without fine-tuning (e.g., bert-base-chinese with a general sentiment head); (c) a simpler machine learning approach (e.g., SVM or logistic regression with TF-IDF features); (d) majority-class or random-guessing baselines for reference. The single reported metric — 0.59 accuracy — is presented in isolation, making it impossible to assess whether the fine-tuning procedure provided any meaningful improvement over simpler alternatives. This is a significant methodological omission, particularly given that the accuracy value itself is relatively modest.

• Generation budget / compute accounting. This paper does not use "generation budget" in the sense common to LLM inference-time compute studies. The computational cost of the pipeline has two components: (1) model fine-tuning on ~2,158 training examples for 5 epochs with batch size 8 — a modest computational requirement by modern standards, likely completable on a single GPU in minutes to hours; (2) inference on ~36,000 posts, which is similarly lightweight. The paper does not discuss computational constraints, does not report training time or hardware specifications, and does not account for the human annotation cost (nine annotators labeling 2,797 posts) in any efficiency metric. The practical scaling limitation is not compute but human labeling effort — expanding the labeled training set or performing error analysis would require additional manual annotation, which the paper does not scale beyond the initial 2,697 labeled posts.

• Cross-validation / statistical protocol. The paper does not employ cross-validation. The labeled data is split once into training and validation sets at a 4:1 ratio, with no mention of k-fold cross-validation, repeated random splits, or held-out test sets. The seed value of 42 is reported for reproducibility, suggesting a single fixed random split was used. There is no reporting of standard deviation across multiple training runs, no confidence intervals on the reported accuracy, and no statistical testing of whether observed differences in sentiment proportions across time periods or provinces are significant. Two-fold cross-validation is mentioned in the template's reference example (the PaLM 2-S* paper on compute-optimal test-time scaling) but is not part of this paper's methodology. Additionally, there is no reported inter-annotator agreement metric (e.g., Cohen's kappa, Fleiss' kappa, or percentage agreement) for the nine human annotators, meaning the reliability of the ground-truth labels against which accuracy is measured cannot be evaluated.

Main Quantitative Results

Temporal Analysis: The July Surge and Sentiment Composition Shift

The central temporal finding is a two-part claim: (1) 89.56% of all Apollo Go-related posts from January–July 2024 are concentrated in July alone, and (2) while sentiment was predominantly positive during the pre-surge period, the proportion of negative comments increased after Apollo Go became a trending topic. These claims are supported by Tables 4 and 5 and visualized in Figures 4–7.

Volume concentration. Table 4 reports the monthly post counts: January (44), February (81), March (113), April (106), May (742), June (685), and July (19,597). The July total represents 19,597 / (44 + 81 + 113 + 106 + 742 + 685 + 19,597) = 19,597 / 21,368 ≈ 91.7% of all posts. The paper reports 89.56% — the small discrepancy may arise from excluding some posts during cleaning or using a slightly different denominator that includes only posts surviving the cleaning process. Either way, the concentration is extreme: the first six months combined contribute only about 8–10% of the dataset.

Sentiment composition before the surge (January–June). Table 4 shows the month-by-month breakdown. Aggregating across January–June: Negative = 9 + 12 + 22 + 19 + 118 + 212 = 392; Neutral = 8 + 39 + 30 + 46 + 190 + 106 = 419; Positive = 27 + 30 + 61 + 41 + 434 + 367 = 960. Total = 392 + 419 + 960 = 1,771. The sentiment proportions are: Positive 54.2%, Neutral 23.7%, Negative 22.1%. The positive-to-negative ratio is approximately 2.45:1 — positive posts outnumber negative posts by more than 2-to-1. This pre-surge period is dominated by Hubei and Guangdong provinces (Table 6: Hubei 383 posts, Guangdong 362), both of which are operational provinces where users have direct experience with Apollo Go.

Sentiment composition during the surge (July). Table 4 reports July totals: Negative 6,552; Neutral 4,670; Positive 8,375. Total = 19,597. Proportions: Positive 42.7%, Neutral 23.8%, Negative 33.4%. The positive-to-negative ratio is approximately 1.28:1 — dramatically compressed from the pre-surge 2.45:1. Negative sentiment has grown from 22.1% of posts to 33.4%.

The daily surge structure (July 1–14). Table 5 and Figure 5b provide daily granularity for the first half of July. The volume trajectory shows a clear escalating pattern:

The data reveal a non-obvious pattern: the surge begins on July 8 with a heavily positive ratio (702 positive vs. 188 negative, 3.73:1), suggesting that initial viral attention was driven by enthusiasm. But as volume peaks on July 11–12, the ratio collapses toward parity — July 12 shows 2,429 positive vs. 2,273 negative (1.07:1). This temporal structure — initial enthusiasm followed by increasingly balanced/negative sentiment as the audience broadens — is the empirical basis for the claim that "negative comments began to rise after it became a hot topic" (Abstract). The paper does not provide sentiment data for July 15–31, so it is unknown whether the negative trend continued or reversed in the latter half of the month. Figure 7 (line chart for July) shows the same data as Table 5 in connected-line form, visually emphasizing the positive-negative convergence around the July 12 peak.

Weekly trend visualization. Figure 4 displays the weekly discussion trend for the full year with emphasis on July (week 27 onward). The figure confirms the pattern visible in Tables 4–5: a flat baseline near zero for January–June, a sharp spike beginning in early July, and a peak around mid-July. The paper notes that data collection extends only through July 14 for the detailed daily analysis, with full-month data shown in the monthly Table 4. Figure 6 and Figure 7 provide line-chart visualizations of sentiment trends for the January–June period and July respectively, with separate lines for positive, neutral, and negative categories.

Geographic shift in discussion volume. Table 6 compares the top 10 provinces by post count for January–June versus July:

The key shift: Hubei drops from 1st to 3rd place while Guangdong and Beijing surge to dominate. The paper notes that "the level of discussion abroad has also surpassed that in Wuhan" (Section 4.2), though it does not provide a separate international post count in Table 6. The July top 10 includes five provinces (Zhejiang, Jiangsu, Shandong, Henan, Fujian) that either do not have Apollo Go operations or have limited pilots — their collective July post count (4,979) exceeds the combined January–June total for all provinces (1,771). This establishes the decoupling of discourse geography from deployment geography during the viral surge.

Spatial Analysis: Province-Level Sentiment Distribution and the Xinjiang-Tibet Divergence

The spatial analysis examines how sentiment toward Apollo Go varies across Chinese provinces, using the IP location metadata inherent in Weibo posts. The key findings are visualized in Figures 9–12.

Aggregate spatial distribution (January–July combined). Figure 9 maps the raw count of negative, neutral, and positive posts by province, using side-by-side choropleths. Figure 10 maps the percentage distribution of sentiment categories within each province, using a tricolor scheme (blue = negative, green = neutral, red = positive). The paper reports that "eastern coastal provinces show a higher level of discussion intensity regarding Apollo Go, particularly in Beijing, Guangdong, and Shandong" (Section 4.2). This pattern is visible in Figure 9 as darker shading (higher total post counts) along the eastern seaboard.

The Xinjiang-Tibet divergence. In Figure 10 (percentage distribution map), the paper identifies a striking geographic pattern: "Xinjiang and Qinghai provinces are predominantly shaded in deep red, indicating optimism about the unstoppable trend of autonomous driving and anticipation for its development. Tibet and Gansu provinces are primarily shaded in deep blue, reflecting concerns about the impact of autonomous vehicles on traditional taxi and ride-hailing services" (Section 4.2). These four provinces share key characteristics — all are in western China, all are non-operational for Apollo Go (no commercial deployment exists in any of them), and all have relatively low total discussion volume — yet they exhibit diametrically opposite sentiment distributions. The paper does not provide the raw sentiment counts or percentages for individual provinces, so the exact magnitude of the divergence cannot be quantified from the reported data.

Correlation with deployment geography. The paper states there is "a strong correlation between the provinces with high discussion intensity and those where Apollo Go operates" (Section 4.2). This is supported by comparing the top provinces in Table 6 with the operational provinces listed in Table 3: Beijing, Hubei, Guangdong, Shanghai, and Zhejiang all appear in both lists. However, the correlation is not perfect — Shandong (ranked 7th in July, 1,051 posts) has no Apollo Go operations according to Table 3, yet generates substantial discussion volume. This suggests that while deployment drives initial discourse, media coverage and national attention can create high discussion intensity in non-operational provinces.

Daily spatial snapshots (July 8–14). Figures 11 and 12 provide day-by-day spatial maps for the peak surge week. Figure 11 (raw counts) shows the geographic expansion of discussion: early days (July 8–9) show concentrated activity in a few eastern provinces, while later days (July 11–13) show more widespread distribution across the country. Figure 12 (percentage distribution) reveals that "overall, there is an expansion in the coverage of positive attitudes and a reduction in the coverage of negative attitudes" over the course of the week (Section 4.2), though the maps themselves are not quantified with specific province-level percentage changes. The paper notes that "provinces with high values of positive attitudes typically align with provinces where there are also high values of neutral attitudes" and attributes this to extensive media coverage "highlighting the benefits of autonomous ride-hailing services such as convenience and technological advancements, thereby mitigating public concerns" (Section 4.2). This claim — that media coverage drives positivity — is interpretive and not directly tested against a measure of media coverage intensity by province.

Geographic findings without quantified baseline. The paper does not provide a statistical test for whether the observed provincial sentiment differences are larger than would be expected by chance given the sample sizes in each province. For provinces with low post counts (which includes most western provinces), the percentage estimates have high variance — a province with 20 total posts and 12 positive posts shows 60% positive, but the 95% confidence interval on that proportion is approximately 36–81%, meaning visual differences between such provinces may not be statistically reliable.

Sentiment Content Analysis: Word Clouds and Thematic Divergence

The word cloud analysis (Figure 13) provides qualitative characterization of what positive and negative commenters discuss. The paper reports:

Positive comment keywords (Figure 13a): "driverless" (无人驾驶), "automatic" (自动), "technique" (技术), "technology" (科技), and "development" (发展). Contextual inspection reveals that "most of the positive comments focused on the discussion of technology applications (such as self-driving and artificial intelligence), personal experience of Apollo Go, as well as the positive news and reports about the self-driving car company" (Section 4.3).

Negative comment keywords (Figure 13b): "response" (响应), "unemployment" (失业), "market" (市场), and "problem" (问题). Contextual inspection reveals that "negative discussions focused mainly on the concerns about job displacement from driverless cars in transportation roles and the safety issues, as well as the negative news such as robotaxi breakdown" (Section 4.3). The paper also notes, from inspecting posts directly, that "some comments reflect dissatisfaction with traditional ride-hailing drivers' behaviors, such as smoking and chatting, suggesting a preference for the perceived comfort and reliability of autonomous services" (Section 5) — a finding that emerges from the positive posts but references a negative comparison target (traditional taxi service).

Thematic asymmetry. The word cloud decomposition reveals that positive and negative sentiment are anchored in qualitatively different domains: positive posts center on technology and personal experience, while negative posts center on socioeconomic threat (unemployment) and operational problems. The paper does not quantify the relative frequency of these themes beyond the visual word cloud representation, so claims about which theme is "dominant" rest on visual inspection rather than systematic frequency comparison.

International comparison (not a quantitative result but a contextual observation). Section 5 notes that "this is similar with studies conducted in the U.S." for positive sentiment themes, but that "in other parts of the world, there is more discussion about personal privacy and data security" — a theme the paper finds "absent" from Chinese Weibo discussions. This observation is not derived from a formal comparative analysis (no non-Chinese social media data is analyzed) but from comparing the paper's findings against the cited literature (Chen and Tomblin, 2021; Jiang and Zheng, 2021).

Ablation Studies and Robustness Checks

This paper does not conduct ablation studies in the conventional machine learning sense. There is no systematic investigation of how model performance changes when architectural components are removed, when hyperparameters are varied, or when different training data subsets are used. The paper's empirical focus is on applying the trained model to produce descriptive analyses, not on evaluating the model's internal workings or validating its design choices. The following are the closest equivalents to robustness checks present in the paper:

Manual annotation quality control. The paper employs nine annotators and provides operational definitions for each label category (Section 2.2). However, no inter-annotator agreement metric is reported — there is no Cohen's kappa, Fleiss' kappa, Krippendorff's alpha, or even simple pairwise percentage agreement. This means the reliability ceiling for model accuracy cannot be established: if annotators disagree on, say, 10% of posts, then a model accuracy of 0.59 should be compared against a human-level ceiling below 1.0, not against perfect classification. The absence of this metric is a significant gap in assessing the training signal quality.

Data cleaning validation. The paper removes official media accounts via username matching and manually annotates 20% of long texts to identify irrelevant posts. The 20% long-text manual annotation could be considered a form of validation for the relevance filtering procedure, but the paper does not report how many of the manually inspected long texts were found to be irrelevant, what feature extraction method was subsequently applied to the remaining 80%, or what accuracy that feature extraction method achieved.

Temporal disaggregation as an implicit robustness check. The paper separates January–June analysis from July analysis and further breaks July into daily snapshots. If the model were producing systematically biased sentiment labels (e.g., consistently misclassifying posts with technical vocabulary as positive), this bias would presumably be uniform across time periods. The observation that sentiment proportions change across time (positive dominant in early July, negative rising toward parity by July 12) suggests that the model is detecting genuine variation rather than simply applying a fixed bias. However, this is not a formal robustness test — it does not rule out the possibility that the content distribution shifted in ways that interact with model bias (e.g., if negative posts in the surge period use different vocabulary than negative posts in the pre-surge period, and the model is more accurate on one vocabulary distribution than the other).

Spatial disaggregation as an implicit robustness check. Similarly, the observation that sentiment varies systematically by province (Xinjiang red, Tibet blue) is difficult to explain purely through model bias unless the model is biased in geographically correlated ways — for instance, if it systematically misclassifies posts from certain provinces due to dialectal or topical vocabulary differences. The paper does not investigate this possibility.

No sensitivity analysis of the accuracy threshold. The model achieves 0.59 accuracy, but the paper does not explore how sensitive the downstream findings are to classification quality. The implicit assumption is that 0.59 is "good enough" for aggregate-level analysis. But if 41% of posts are misclassified, and if misclassification is not symmetric across categories (e.g., the model might be more likely to confuse neutral with positive than neutral with negative, creating systematic inflation of positive counts), the aggregate sentiment proportions could be biased in ways that affect the main findings. The paper does not report a confusion matrix, so the direction and magnitude of any such bias cannot be evaluated.

No alternative model comparison. The paper does not compare the fine-tuned BERT model against any simpler baseline. If a lexicon-based sentiment scorer with a Chinese sentiment dictionary achieved similar accuracy on the held-out set, the case for using BERT specifically (rather than a simpler, more interpretable method) would be weaker. If the fine-tuned BERT substantially outperformed simpler methods, that would strengthen the methodological contribution. Neither comparison is made.

No training set size sensitivity. The 2,697 labeled posts (4:1 train/validation split) represent a modest training set. The paper does not investigate whether accuracy would improve with more labeled data (e.g., by plotting a learning curve), whether certain categories benefit more from additional examples than others, or whether the 0.59 accuracy represents a plateau or could be improved with further annotation effort. This is particularly relevant because the paper positions its methodology as reproducible — future researchers applying similar approaches would benefit from knowing how much labeling effort is needed to achieve adequate performance.

Critical Assessment

Does the study demonstrate that public sentiment toward Apollo Go can be measured from Weibo data using fine-tuned BERT?

The paper demonstrates that a fine-tuned BERT model can be applied to Weibo posts to produce sentiment labels, but the strength of this demonstration is limited by the modest accuracy (0.59) and the absence of baseline comparisons. An accuracy of 0.59 means that on the validation set, roughly 4 out of every 10 classifications are incorrect. This is substantially better than random guessing for three classes (~0.33), but the magnitude of the improvement over random guessing is modest — the model is getting about 2.6 additional correct classifications per 10 posts compared to random chance. Without a comparison against simpler methods (lexicon-based, SVM, or even the unfine-tuned bert-base-chinese with a general sentiment head), it is impossible to know whether the fine-tuning on domain-specific data provided meaningful benefit or whether comparable performance could have been achieved with dramatically less effort.

The claim that the pipeline successfully measures public sentiment is therefore supported but with significant caveats. The model does extract signal above chance level, and the temporal and spatial patterns it reveals (the July surge, the geographic variation, the positive-negative thematic divergence) are directionally plausible and consistent with external context (Apollo Go did go viral in July 2024, and the themes of technology enthusiasm vs. labor anxiety are well-documented in the AV acceptance literature). But the paper does not quantify how much noise the 41% misclassification rate introduces into the reported sentiment proportions, and the absence of confidence intervals or statistical tests means that patterns observed in the data cannot be distinguished from sampling variation or systematic classification bias.

Does the study demonstrate that negative sentiment increased after Apollo Go became a trending topic?

This claim is supported by the raw counts in Tables 4 and 5, which show the positive-to-negative ratio declining from ~2.45:1 in January–June to ~1.28:1 in July, and further compressing toward ~1.07:1 on the peak day (July 12). The temporal pattern is internally consistent and shows a clear directional trend.

However, the causal interpretation — that going viral caused the increase in negative sentiment proportion — is not directly tested. The mechanism could be compositional rather than causal: the July surge brought a different population of discussants (national audience, media consumers, people without direct Apollo Go experience) into the conversation, and this population had a systematically different baseline sentiment distribution than the early-adopter population that dominated January–June discussion. The negative sentiment proportion increased not because individuals changed their minds, but because the sample composition changed. The paper's spatial analysis indirectly supports this compositional interpretation: the July discussion geography expanded to include many non-operational provinces, and the geographic shift in who was talking likely drove the shift in what was being said. The paper acknowledges this implicitly when noting that Hubei discussion may have "become localized and primarily happening offline" (Section 4.2) — the early adopters didn't become more negative; they were simply numerically swamped by a larger, different audience.

The claim is therefore better characterized as supported as a descriptive finding (the proportion of negative posts increased) but untested as a causal claim about sentiment change. The paper's narrative framing (negative comments "began to rise after it became a hot topic") could be read as implying that the trending status caused more negative sentiment, but the data only support the weaker claim that trending status coincided with a different sentiment composition.

Does the study demonstrate that public sentiment varies significantly by province, with non-operational provinces showing divergent attitudes?

This is the paper's most robust finding, but it also has the weakest quantitative support in terms of reported evidence. The maps in Figures 9 and 10 visually demonstrate geographic heterogeneity — different provinces have different color distributions, and the Xinjiang-Tibet contrast is visually striking. However, the paper does not provide:

• Province-specific sentiment counts or percentages in tabular form. For example, what is the exact positive percentage for Xinjiang vs. Tibet? How many posts is each estimate based on? The claim that Xinjiang shows "optimism" and Tibet shows "concern" rests on visual inspection of a choropleth map rather than reported numerical comparisons.
• Statistical tests of geographic variation. Are the inter-province differences larger than would be expected under random sampling variation? A province with 30 total posts has a much wider confidence interval on its sentiment proportion than a province with 3,000 posts, meaning visual differences between low-volume provinces could arise by chance.
• Control for confounders. Is the Xinjiang-Tibet divergence driven by a handful of highly active users in each province whose individual sentiments dominate the aggregate? Weibo posts are not independent observations (one user can post many times), but the paper does not analyze user-level clustering.

The spatial heterogeneity claim is credible and visually supported but falls short of the evidentiary standards needed for strong quantitative conclusions. The finding is better characterized as an exploratory observation that merits follow-up with larger samples, statistical testing, and control for user-level clustering, rather than as a definitively established pattern.

Does the study demonstrate that positive and negative sentiment are driven by fundamentally different themes (technology vs. labor anxiety)?

The word cloud analysis in Figure 13 provides direct visual evidence for thematic asymmetry: "technology" and "driverless" dominate the positive cloud, while "unemployment" and "market" dominate the negative cloud. The paper supplements this with qualitative post inspection that confirms the word-level patterns reflect coherent thematic clusters.

This finding is supported as a qualitative characterization of the discourse. The word cloud methodology is inherently limited — it captures unigram frequency without accounting for negation, sarcasm, or multi-word phrases, and it does not quantify the proportion of positive posts that discuss technology vs. other positive themes. But within those limitations, the thematic divergence is clear and consistent with the broader AV acceptance literature, which has long identified technology enthusiasm and labor displacement as distinct attitudinal dimensions. The paper's contribution is not discovering that these themes exist (survey research has documented them) but rather showing that they emerge organically as the dominant clustering dimensions in spontaneous Chinese social media discourse about an operational robotaxi service.

A stronger demonstration would have involved topic modeling (e.g., LDA) to quantify the prevalence of each theme, or aspect-based sentiment analysis to identify which specific aspects of Apollo Go (technology, price, safety, labor impact) attract positive vs. negative sentiment. The word cloud approach provides a useful overview but leaves most of the analytical potential of the text data unexploited.

What experiments would have strengthened the paper?

Several additional analyses would have substantially increased the paper's empirical rigor:

1. A confusion matrix for the fine-tuned model, showing which categories are confused with which. If the model frequently confuses neutral with positive but rarely confuses positive with negative, the estimated positive proportion is likely inflated — a critical issue for the temporal trend analysis.

2. Baseline model comparisons: accuracy of the fine-tuned BERT vs. (a) a Chinese sentiment dictionary, (b) an unfine-tuned general-purpose Chinese sentiment classifier, (c) a simple bag-of-words logistic regression or SVM. This would contextualize the 0.59 accuracy figure and justify the use of BERT specifically.

3. Inter-annotator agreement metrics: Cohen's kappa or Fleiss' kappa for the nine annotators. This would establish the reliability ceiling — if human agreement is, say, 0.70, then a model accuracy of 0.59 is much more impressive than if human agreement is 0.95.

4. Province-level sentiment tables with sample sizes: replacing the purely visual choropleth maps with a supplemental table listing the number of posts, positive/negative/neutral counts, and proportions for each province. This would allow readers to assess whether the Xinjiang-Tibet divergence is statistically meaningful given the sample sizes.

5. User-level analysis: the data contains User ID fields. The paper could report how many unique users contribute to each province's discussion, whether a small number of prolific users drive the aggregate sentiment in certain provinces, and whether sentiment varies systematically between high-volume and low-volume posters.

6. Temporal analysis beyond July 14: the daily breakdown stops at July 14 (Table 5), and the full monthly data (Table 4) shows July ending with 19,597 posts. What happened in the second half of July? Did the negative trend continue, reverse, or plateau? Extending the daily analysis to July 31 would provide a more complete picture of the surge-decay cycle.

7. Topic proportions over time: rather than static word clouds, tracking how the prevalence of key themes (technology, unemployment, safety, price) changed from January–June to July would directly test the claim that the thematic composition of discourse shifted with the audience expansion.

8. Sentiment accuracy calibration by text length: the paper notes that long texts posed annotation challenges. If model accuracy varies with post length — performing better on short, straightforward posts and worse on long, ambiguous ones — the aggregate sentiment estimates could be biased toward the sentiment expressed in short posts.

6. Limitations and Trade-offs

Limitation 1: Model Accuracy of 0.59 Is Not Benchmarked and Its Downstream Consequences Are Unquantified

The assumption or constraint. The paper's entire analytical edifice — temporal trends, spatial distributions, thematic characterization — rests on the sentiment labels produced by the fine-tuned BERT model. The model achieves a reported accuracy of 0.59 on the held-out validation set (Section 3). The paper acknowledges this figure without qualification or comparison: no baseline accuracy is provided for a majority-class classifier, a lexicon-based approach, or an unfine-tuned pretrained sentiment model, and no per-class precision, recall, or confusion matrix is reported. Section 5 (Conclusion and Discussion) does not mention model accuracy among the acknowledged limitations.

The consequence. An accuracy of 0.59 means that approximately 41% of the model's sentiment classifications are expected to be incorrect. This has direct, quantifiable consequences for every downstream finding. Consider the temporal analysis: the paper reports that positive-to-negative sentiment ratio declined from ~2.45:1 in January–June to ~1.28:1 in July, with a further compression to ~1.07:1 on July 12 (Tables 4 and 5). If misclassifications are asymmetric — for instance, if the model systematically confuses neutral posts with positive posts more often than with negative posts — the estimated positive proportions in both time periods would be inflated, and the magnitude of the July decline would be distorted. The direction of distortion cannot be assessed because no confusion matrix is reported. The spatial findings face an even steeper problem: for low-volume provinces like Xinjiang and Tibet (which the paper highlights as showing diametrically opposite sentiment), the sample sizes are presumably small, and a 41% misclassification rate means that individual province-level sentiment percentages are extremely noisy. A province with 40 total posts and 25 labeled positive (62.5% positive) could have a true positive proportion anywhere in a wide confidence interval if 41% of those 25 "positive" labels are actually misclassified neutrals or negatives.

What evidence exists in the paper. The single accuracy figure (0.59) is reported in Section 3 without further decomposition. The paper does not report a confusion matrix, per-class F1 scores, precision-recall curves, or any form of error analysis. There is no investigation of whether misclassifications are symmetric or systematic, whether accuracy varies by post length or topic, or whether certain sentiment categories are more reliably classified than others. The validation set is described as a 4:1 split of the 2,697 labeled posts (~539 validation samples), but the class distribution of the validation set is not reported, so it is unknown whether the 0.59 accuracy reflects balanced performance or degenerate behavior on minority classes. There is no baseline comparison — no accuracy figure for a random classifier, a majority-class classifier, a lexicon-based sentiment scorer, or an unfine-tuned model — making it impossible to assess whether the fine-tuning procedure provided meaningful improvement over trivial alternatives.

Mitigation status. The paper does not attempt to address this limitation. Section 5 (Conclusion and Discussion) lists three limitations: lack of demographic information, small data volume and time span, and data concentration in a single platform. Model accuracy is not mentioned as a limitation, and no future work is proposed to improve classification performance, characterize error patterns, or quantify how classification noise propagates into downstream findings. The paper implicitly treats the 0.59 accuracy as sufficient for the aggregate analyses presented, but this assumption is never defended or tested.

Limitation 2: The Difficulty Estimation Cost Is Externalized — 89.56% of Data Comes from a Single Month

The assumption or constraint. The paper's temporal analysis reveals that 89.56% of all Apollo Go-related posts from January–July 2024 are concentrated in July alone (Section 4.1, Table 4), with the January–June period contributing only 1,771 posts total across six months. The paper structures its analysis around the contrast between these two periods: January–June as the "pre-trending" baseline and July as the "post-trending" surge. The paper acknowledges that "the data volume and time span are relatively small, with discussions on social media platforms heating up late in the study period" and notes that "other similar studies typically observe data continuously for about three months, offering a more comprehensive view of public sentiment over time" (Section 5).

The consequence. The extreme temporal concentration creates a fundamental asymmetry in the reliability of the temporal comparison that is the paper's central finding. The pre-trending "baseline" sentiment distribution (January–June: 54.2% positive, 23.7% neutral, 22.1% negative, computed from Table 4) is estimated from 1,771 total posts — a sample that is more than an order of magnitude smaller than the July sample (19,597 posts). Moreover, the pre-July sample is not a stable baseline but is itself trending upward: January had 44 posts, February 81, March 113, April 106, May 742, and June 685. The composition of discussants was likely already shifting during May and June (when volume increased ~7× relative to the prior months), meaning the "pre-trending" period is not a uniform baseline but rather an early ramp-up phase whose sentiment distribution may already reflect the beginnings of national attention rather than a pure "early adopter" signal. The paper claims that "negative comments began to rise after it became a hot topic on July 21" (Abstract), but the comparison is between a highly aggregated six-month baseline with heterogeneous internal dynamics and a single-month surge period — the temporal resolution is too coarse, and the pre-surge sample too small, to isolate a specific inflection point at July 21 with any precision.

Furthermore, the study only captures the onset of the viral surge, not its aftermath. The daily analysis in Table 5 covers only July 1–14 — just the first half of the surge month. The paper reports July totals of 19,597 posts in Table 4, but only 14,469 of these are accounted for in the July 1–14 daily breakdown. What happened to the remaining ~5,000 posts from July 15–31? Did the negative sentiment trend continue to rise, plateau, or reverse? The paper cannot say, because the data for the latter half of July is not analyzed at daily granularity. This means the study captures the acceleration of negative sentiment during the peak of the surge but cannot characterize the trajectory — whether the sentiment shift was a temporary reaction to the viral moment that subsided as attention waned, or a permanent recalibration of public attitudes. The practical implication for a policymaker is ambiguous: should they expect the negative sentiment spike to be a transient social media phenomenon or a durable shift in public perception?

What evidence exists in the paper. The monthly breakdown in Table 4 reveals the extreme concentration. The daily breakdown in Table 5 reveals that the analysis truncates on July 14. The paper states in Section 5 that "the data volume and time span are relatively small" and that extending the observation period would offer "a more comprehensive view of public sentiment over time," acknowledging this limitation explicitly.

Mitigation status. Partially acknowledged but not addressed. The paper recognizes the short time span as a limitation in Section 5 but does not conduct sensitivity analysis — for instance, comparing sentiment trends in the subset of July data that corresponds to the same volume as the entire January–June period (roughly the first ~1,700 July posts) to see whether the "negative increase" pattern holds when sample sizes are equalized. The paper recommends future work to "extend the observation period for a more comprehensive analysis" but does not report any post-July data that might already exist or describe plans for continued monitoring.

Limitation 3: Single Platform, Single Service, Single Language — The Findings Cannot Be Assumed to Generalize

The assumption or constraint. All data comes from a single social media platform (Weibo) discussing a single autonomous ride-hailing service (Baidu Apollo Go) in a single language (Chinese) within a single national context (China). The paper acknowledges this implicitly when it notes that findings are "similar with studies conducted in the U.S." for positive sentiment themes but that privacy concerns — prominent in Western discussions — are absent from the Chinese Weibo discourse (Section 5). The paper also notes that some negative Chinese comments focus on "dissatisfaction with traditional ride-hailing drivers' behaviors, such as smoking and chatting" — a theme that may reflect local service quality dynamics rather than universal attitudes toward autonomous vehicles. The paper positions its scope as China-specific, stating that Apollo Go is "China's first shared autonomous driving service platform operating in multiple cities" (Section 1).

The consequence. Three distinct generalization failures are possible, and none are tested. First, platform-specific user demographics: Weibo's user base is not demographically representative of the Chinese population, and the subset that posts about Apollo Go is further self-selected. Users who post about autonomous vehicles on social media are likely younger, more urban, more tech-savvy, and more opinionated than the general population. The sentiment distribution on Weibo may overrepresent technology enthusiasm (because tech-interested users are more likely to post) or overrepresent labor anxiety (because those threatened by automation are motivated to voice opposition), and without demographic data (which the paper acknowledges it lacks), the direction and magnitude of platform bias cannot be assessed. Second, service-specific dynamics: Apollo Go's pricing, deployment strategy, marketing, and incident history are unique. If Apollo Go experienced a specific negative event — a well-publicized safety incident, a controversial pricing change, a labor dispute — during the study period, the sentiment patterns would reflect Apollo Go's management rather than public attitudes toward autonomous ride-hailing in general. The paper does not control for or document Apollo Go-specific events beyond the Apollo Day 2024 event (May 15). Third, cultural specificity: the paper's finding that Chinese Weibo discourse lacks privacy and data security concerns (unlike Western AV discussions) suggests that the thematic structure of sentiment is culturally conditioned. The dominance of labor displacement as a negative theme may reflect specific features of China's labor market, media environment, or regulatory context that do not transfer to other countries. A transportation agency in Germany or the United States cannot assume that the sentiment clusters identified in this study (technology enthusiasm vs. unemployment anxiety) characterize their own public's concerns.

What evidence exists in the paper. No evidence is presented to support generalization claims because the paper does not attempt cross-platform, cross-service, or cross-national validation. The paper does note international differences: Section 5 states that "in other parts of the world, there is more discussion about personal privacy and data security," citing Chen and Tomblin (2021), and attributes this to "the longer timeframe for the launch of self-driving services in those regions, allowing for a broader range of issues to surface." This is a post-hoc interpretation, not a tested hypothesis. The paper does not analyze non-Weibo Chinese social media (e.g., Zhihu, Douyin, Xiaohongshu), does not compare Apollo Go sentiment to sentiment about competing Chinese autonomous ride-hailing services (e.g., Pony.ai, AutoX, WeRide), and does not apply the same methodology to discussions of Waymo or Cruise on English-language Twitter for comparative analysis.

Mitigation status. Minimally acknowledged. The paper recognizes in Section 5 that comparing its findings with other studies requires caution because of different timeframes and cultural contexts, but it does not label the single-platform, single-service, single-national-context scope as a limitation that requires explicit caveats on all findings. The future work recommended in Section 5 — exploring "demographic characteristics on public attitudes" and extending "the observation period" — addresses other limitations but not the fundamental scope constraint. The paper does not suggest multi-platform, multi-service, or cross-national comparative analysis as a direction for future work.

Limitation 4: No Demographic Data — The Study Cannot Explain Who Holds Which Attitudes or Why

The assumption or constraint. Weibo's API provides user metadata including User ID, IP location (to province level), and engagement metrics, but it does not expose demographic information such as age, gender, income, education, occupation, or urban/rural residence within a province. The paper explicitly states this limitation: "the lack of social demographic information, such as age and income, limits our understanding of the factors influencing public attitudes" (Section 5), and notes that "existing studies have shown that these attributes are critical in shaping public perceptions" (citing Wang et al., 2022, on impacting factors for public attitudes toward AVs using sentiment analysis from social media data).

The consequence. The paper's findings are entirely descriptive — they characterize what the sentiment distribution looks like across time and space but cannot explain why it looks that way. This limitation cascades through every major finding:

• The Xinjiang-Tibet divergence (Section 4.2, Figures 9–10) is the paper's most striking spatial finding, but without demographic data, the paper cannot test competing explanations. Is Xinjiang's optimism driven by a younger population, different industrial composition (less reliance on taxi/driving jobs), different media consumption patterns, or a different political/cultural orientation toward technological modernization? Is Tibet's concern driven by an older population, higher dependence on traditional transportation employment, or exposure to different news coverage? The paper can only note the divergence exists — it cannot advance understanding of its causes.

• The negative sentiment increase in July (Section 4.1) is attributed to the viral surge bringing a broader audience into the conversation, but without demographic data, this compositional explanation cannot be tested. Did the July surge bring in users who are demographically different from the January–June discussants (e.g., older, less urban, lower income, different occupational sectors), or did it amplify the same demographic groups? If the former, the sentiment shift is a composition effect and does not reflect attitudinal change within any group; if the latter, it reflects genuine opinion dynamics among a stable demographic. The paper cannot distinguish these possibilities.

• The word cloud thematic analysis (Section 4.3, Figure 13) identifies unemployment as a dominant negative theme and technology enthusiasm as a dominant positive theme, but cannot identify who is concerned about unemployment. Are negative posts about job displacement concentrated among users in transportation-sector-heavy provinces? Among older users who fear technological obsolescence? Among lower-income users who perceive autonomous vehicles as a threat to accessible employment? Without demographics, the paper can report what themes appear but cannot connect themes to populations, which is what a policymaker designing targeted interventions (e.g., workforce retraining programs for taxi drivers) would need to know.

• The "dissatisfaction with traditional taxi drivers' behaviors" finding (Section 5) is suggestive but cannot be contextualized: are these complaints coming from young, urban, high-income users who have alternatives to taxis? From users with disabilities or mobility constraints for whom traditional taxi driver behavior has been a barrier? Or from users who simply prefer the consistent, impersonal experience of an autonomous vehicle? The paper cannot say, and so the finding remains an isolated observation rather than an actionable insight.

What evidence exists in the paper. The paper explicitly acknowledges this limitation in Section 5, stating that "future research to explore the impact of various demographic characteristics on public attitudes" would be beneficial and that the current study lacks demographic information that prior research has shown to be important (citing Wang et al., 2022). However, the paper does not attempt to mitigate this limitation with proxy demographic indicators — for example, correlating provincial sentiment with provincial aggregate statistics (average income, age distribution, employment in transportation sectors, urbanization rate) from Chinese statistical yearbooks, which could provide suggestive evidence about demographic drivers even in the absence of individual-level data.

Mitigation status. Acknowledged but not addressed. The limitation is listed among three in Section 5 alongside data volume and time span constraints, and future work is proposed. But the paper does not attempt any form of ecological inference or aggregate demographic correlation that might partially mitigate the limitation with available data. The absence of demographic analysis is particularly consequential because the paper's core contribution — spatial heterogeneity in sentiment — is exactly where demographic data would be most illuminating: explaining why Xinjiang differs from Tibet, rather than merely noting that a difference exists.

Limitation 5: No Inter-Annotator Agreement Metrics — The Ground Truth Labels Have Unknown Reliability

The assumption or constraint. The fine-tuned BERT model is trained and evaluated on 2,697 posts manually labeled by nine annotators using the doccano tool (Section 2.2). The paper provides operational definitions for each category — positive (excitement, enthusiasm, support, encouragement), negative (dissatisfaction, disgust, disapproval, abusive language), neutral (objective facts without emotional bias), and drop (irrelevant content) — but it does not report any measure of inter-annotator agreement. There is no Cohen's kappa, Fleiss' kappa, Krippendorff's alpha, percentage agreement, or any other reliability metric for the annotation process. The model's reported accuracy of 0.59 is computed against these labels without any estimate of the labels' reliability ceiling.

The consequence. Without inter-annotator agreement metrics, the quality of the training signal is unknown, and the interpretation of the 0.59 accuracy figure is ambiguous in ways that directly affect the strength of every downstream claim. There are two scenarios, with very different implications:

• Scenario A: High annotator agreement (e.g., Fleiss' kappa > 0.8). This would mean the sentiment categories are well-defined for this domain, the annotators consistently applied them, and the labels represent a reliable ground truth. In this scenario, the model's 0.59 accuracy is genuinely modest — the model fails to capture a substantial portion of the signal that human annotators reliably detect. The downstream sentiment estimates contain ~41% noise relative to what a perfect classifier (or a human) would produce.

• Scenario B: Low annotator agreement (e.g., Fleiss' kappa < 0.5). This would mean the sentiment categories are inherently ambiguous for Weibo posts about Apollo Go, different annotators interpret the same post differently, and the "ground truth" labels are themselves noisy. In this scenario, the model's 0.59 accuracy may be approaching the human reliability ceiling — if humans only agree with each other 65% of the time, a model that achieves 59% accuracy is capturing most of the learnable signal. The downstream sentiment estimates would still be noisy, but the noise would originate in task ambiguity rather than model deficiency.

The paper provides no basis for distinguishing between these scenarios. The nine annotators may have achieved high agreement (suggesting the 0.59 model accuracy is a genuine limitation) or low agreement (suggesting the sentiment categories are poorly defined for this data, and that even perfect human annotation would produce an uncertain ground truth). The downstream consequence is that the reliability of every sentiment proportion reported — the 54.2% positive rate in January–June, the 33.4% negative rate in July, the deep-red coloring of Xinjiang in Figure 10 — cannot be assessed relative to the best possible human performance. If human annotators would have produced a different sentiment distribution on the same posts (because the categories are ambiguous), then the model's output, even at higher accuracy, would not converge to a single "true" distribution — and the patterns reported in the paper would depend on the specific operationalization of sentiment categories used by this annotation team.

Furthermore, the annotation process itself has unexamined structure. The nine annotators are not described in terms of expertise or background — were they transportation researchers, NLP specialists, or general crowd-workers? Did they receive training on the category definitions, and was there a calibration process? Were posts assigned to single annotators or multiple annotators? If each post was labeled by only one annotator, individual annotator biases (e.g., one annotator systematically categorizing ambiguous technology-discussion posts as positive) would be baked into the training labels as "ground truth" rather than captured as annotator variance.

What evidence exists in the paper. The annotation process is described in Section 2.2: nine individuals, doccano tool, 2,797 posts annotated, 2,697 retained after removing excessively long and meaningless texts, four categories with operational definitions provided. No inter-annotator agreement metrics are reported, no information about annotator training or calibration is provided, and no information about whether posts received single or multiple annotations is given.

Mitigation status. Not addressed and not acknowledged as a limitation. Section 5 lists three limitations — lack of demographics, small data volume and time span, and potential benefits of combining social media with traditional data — but the absence of annotation reliability metrics is not mentioned. The paper treats the human labels as a gold standard and reports model accuracy against them without interrogating the standard's own reliability. This is a significant methodological gap: establishing annotation quality is a standard practice in NLP research, and its absence weakens the interpretability of the central quantitative metric (0.59 accuracy) and all downstream analyses that depend on the labeled data.

Limitation 6: The Sentiment Taxonomy Collapses Ambivalence — Users with Mixed Feelings Are Forced into a Single Category

The assumption or constraint. The annotation scheme and model classify each Weibo post into exactly one of three sentiment categories: positive, negative, or neutral (with a fourth "drop" category for irrelevant posts). This is a standard sentiment analysis design, but it imposes a simplifying assumption: each post expresses a single, coherent attitudinal stance toward Apollo Go. The operational definitions reinforce this: positive posts "indicate support and anticipation... with the belief that further deep testing and early commercial operation are warranted," while negative posts "indicate a negative attitude... suggesting that the development of this technology should not continue at this time" (Section 2.2). These definitions implicitly assume that a post's author has resolved their attitude into a unidimensional pro- or anti- stance.

The consequence. The paper's own findings undermine this assumption. The word cloud analysis (Section 4.3, Figure 13) reveals that positive and negative sentiment are driven by different thematic concerns — technology enthusiasm vs. labor anxiety. This is not merely a finding about the aggregate discourse; it implies that a single individual could simultaneously hold both attitudes: excitement about autonomous driving technology (a positive-valenced cognition) and concern about job displacement for taxi drivers (a negative-valenced cognition). A post that expresses both — "The self-driving technology is incredible, but I worry about taxi drivers losing their jobs" — must be forced into a single category (positive, negative, or neutral) by the classifier, even though the attitudinal content is genuinely mixed. The paper's finding that some positive posts express dissatisfaction with traditional taxi drivers' behavior (smoking, complaining about short trips) while supporting Apollo Go (Section 5) further demonstrates attitudinal complexity: these posts are positive toward Apollo Go but motivated in part by negative sentiment toward an incumbent service.

The forced-choice classification may systematically misrepresent ambivalent or complex posts in ways that affect the temporal and spatial findings. Consider the temporal shift from January–June to July. The paper finds that negative sentiment proportion increased during the surge. But what if the surge brought in users with more nuanced, ambivalent attitudes — people who had not been following Apollo Go closely before and, when exposed to the national discussion, formed more complex opinions than the early adopters' straightforward enthusiasm? A post that says "This technology is exciting for the future, but the impact on employment is deeply concerning" could be classified as neutral (because it balances positive and negative), positive (if the classifier weights the excitement more heavily), or negative (if it weights the concern more heavily). If such ambivalent posts increased in July and the classifier systematically assigned them to the negative category (because negative keywords like "unemployment" are prominent), the reported increase in negative sentiment would partly reflect an increase in attitudinal complexity rather than an increase in outright opposition. The paper cannot distinguish these possibilities because the classification scheme does not allow for mixed sentiment.

Similarly, the spatial divergence between Xinjiang (optimistic) and Tibet (concerned) could partly reflect differences in how ambivalence is expressed in those provinces' Weibo discourse, rather than differences in underlying attitudes. If Xinjiang users tend to post about technology even when they have concerns, while Tibet users tend to post about concerns even when they are excited about the technology, the forced-category classification would exaggerate the apparent attitudinal gap.

What evidence exists in the paper. The word cloud analysis (Figure 13) and the qualitative finding about dissatisfaction with traditional taxi drivers (Section 5) are the primary evidence that attitudinal content is not unidimensional. The paper does not analyze the co-occurrence of positive and negative themes within the same post, does not compute sentiment intensity scores, and does not attempt multi-label classification (where a post could be tagged as both "technology-positive" and "labor-negative"). The validation set accuracy of 0.59 may partly reflect annotator difficulty in classifying posts that contain mixed sentiment — if annotators disagreed on how to categorize ambivalent posts, this would manifest as lower inter-annotator agreement and a lower model accuracy ceiling, but without agreement metrics, this cannot be tested.

Mitigation status. Not addressed and not acknowledged as a limitation. The paper adopts a standard three-way sentiment classification without discussing its appropriateness for a domain where attitudinal complexity is documented. The word cloud finding that positive and negative themes are distinct (Section 4.3) is presented as a result rather than recognized as evidence that the classification scheme itself may be misspecified for capturing the full structure of public sentiment. Future work on aspect-based sentiment analysis — which would allow separate sentiment scores for different aspects of Apollo Go (technology, price, safety, labor impact) — is not proposed. The paper could have partially mitigated this by analyzing whether posts that contain both positive and negative keywords are more common in certain provinces or time periods, providing a window into ambivalence even within the single-label framework, but this analysis was not conducted.

7. Implications and Future Directions

How This Work Changes the Landscape

This paper does not claim to advance NLP methodology — it deploys a standard fine-tuned BERT classifier with modest accuracy (0.59) — and so its contribution to the sentiment analysis literature itself is incremental rather than transformative. Where the paper does shift the landscape is in the transportation research and urban policy communities, by providing a concrete, reproducible demonstration that social media data with verified geographic metadata (Weibo's IP location feature) can characterize public sentiment toward an operational robotaxi deployment at subnational spatial resolution during a period of rapid public attention, revealing patterns that survey-based research would either miss entirely or capture at much lower temporal and spatial granularity.

This is more than a "social media can be useful for transportation research" observation — that claim has been made before (Ding et al., 2021; Jiang and Zheng, 2021; Chen and Tomblin, 2021; Tori et al., 2024). The paper's specific contribution is establishing that the geographic structure of sentiment toward autonomous ride-hailing does not simply track deployment geography. The paper's most striking spatial finding — that Xinjiang and Qinghai show predominantly positive sentiment while neighboring Tibet and Gansu show predominantly negative sentiment, despite none of these four provinces having Apollo Go operations (Section 4.2, Figures 9–10) — is a genuinely counterintuitive result that challenges the intuitive model in which exposure to autonomous vehicles (through local deployment) drives sentiment formation. This finding implies that sentiment in non-operational regions is shaped by second-order mediators — media framing, economic anxiety about specific industries, cultural attitudes toward technological change, or province-level labor market structures — whose influence can produce diametrically opposite aggregate sentiment in demographically and geographically similar regions.

For transportation policymakers and service providers, this finding reframes the problem of public acceptance. The standard strategy of demonstrating safety and reliability through pilot deployments in specific cities assumes that positive sentiment will diffuse geographically from operational regions outward. The Apollo Go data suggests a more complex reality: when a service goes viral nationally, the online discourse is dominated by people who cannot use it and whose sentiment is shaped by media representation rather than experience. The paper documents that in July 2024, the geography of discussion decoupled from deployment geography — Hubei (where Apollo Go actually operates) dropped from 1st to 3rd place in discussion volume, while Guangdong (3,615 posts), Beijing (2,483 posts), and non-operational coastal provinces like Zhejiang, Jiangsu, and Shandong generated far more discussion than they had in the pre-surge period (Table 6). If the sentiment of these non-user discussants differs systematically from actual users — and the paper's data suggests it does, with the positive-to-negative ratio compressing from ~2.45:1 pre-surge to ~1.28:1 during the surge — then a transportation agency monitoring "public sentiment" through social media volume is measuring a signal increasingly dominated by people with no direct experience of the service.

This is not merely an academic observation; it has direct implications for how cities evaluate whether to permit or expand robotaxi services. If a city council in a non-operational province sees social media data showing rising negative sentiment, they may interpret it as evidence of public opposition — when in fact the negative sentiment is being generated disproportionately by people in provinces where the service doesn't exist, while actual users in Wuhan may be substantially more positive. The paper does not solve this problem (it lacks the data to separate user sentiment from non-user sentiment within each province), but it identifies the problem's existence and provides a template for diagnosing it through geographic disaggregation of social media data.

Reconciling prior contradictions. The paper also contributes to resolving an apparent tension in the autonomous vehicle acceptance literature. Prior survey research (Tussyadiah et al., 2017; Bansal et al., 2016) documented that AV attitudes are multidimensional — people simultaneously hold technology enthusiasm and safety/labor concerns. But social media sentiment analyses, including this one, force posts into single-category classifications (positive, negative, neutral), potentially obscuring this multidimensionality. The paper's word cloud analysis (Section 4.3, Figure 13) helps reconcile these approaches by showing that the aggregate positive and negative discourse clusters around different themes — technology experience vs. unemployment — which is exactly what the survey literature's multidimensional models would predict. The paper demonstrates that even within a forced-choice classification scheme, the thematic divergence between positive and negative posts recovers the underlying attitudinal structure that surveys measure through Likert scales on separate dimensions. This is a methodological bridge: it shows that social media sentiment analysis, despite its coarseness, can recover the same attitudinal dimensions that more expensive survey methods identify, provided the analysis goes beyond simple positive/negative proportions to examine what people are talking about within each category.

Research directions this work makes more attractive. The paper's most significant effect on the research landscape may be to make geographically disaggregated social media analysis of mobility service sentiment a clearly viable methodology for transportation researchers who lack NLP expertise. By releasing the fine-tuned model on Hugging Face and the full code pipeline on GitHub, the paper lowers the barrier to entry for replication and adaptation. A transportation researcher studying, say, public response to a new bike-share program or congestion pricing policy in a Chinese city could adapt this pipeline — same model, same annotation protocol, same spatial aggregation framework — with modest domain-specific fine-tuning effort. This is not a deep NLP contribution, but it is a practical infrastructure contribution to the transportation research community.

Conversely, the paper also suggests that certain research directions are less promising than they might appear. The modest 0.59 accuracy achieved on this domain-specific sentiment task, despite using a pretrained 110M-parameter transformer and manual annotation, implies that off-the-shelf sentiment tools applied to Chinese social media discourse about emerging technologies will likely produce unreliable results without substantial domain adaptation. The paper does not benchmark against lexicon-based methods, but the complexity of the sentiment expressions — posts that mix technology enthusiasm with labor anxiety, posts that express support for Apollo Go through complaints about traditional taxi drivers — suggests that keyword-based approaches would perform substantially worse than the already-modest BERT accuracy. This implies that transportation researchers seeking to use social media sentiment analysis as a measurement tool should budget for domain-specific annotation effort; zero-shot or lexicon-based approaches are unlikely to be adequate for nuanced policy-relevant sentiment measurement.

Magnitude of the shift. This is an incremental empirical contribution with practical consequences rather than a paradigm shift. The paper does not introduce new theoretical frameworks, does not develop new NLP architectures, and does not claim universal findings. Its contribution is demonstrating — with a concrete, reproducible pipeline applied to a novel dataset capturing a live technological deployment — that spatial and temporal sentiment heterogeneity toward robotaxis exists at subnational scale, that this heterogeneity is not reducible to deployment geography, and that the standard tools of fine-tuned BERT plus human annotation can capture it adequately for policy-relevant analysis. For a field (transportation public acceptance research) that has historically relied on surveys with sample sizes in the hundreds, the demonstration that 36,096 social media posts can be classified and mapped to reveal province-level sentiment patterns during a viral deployment event represents a meaningful expansion of the methodological toolkit — not a replacement for surveys, but a complementary source of evidence with different strengths (scale, geographic granularity, temporal resolution) and weaknesses (unknown demographics, platform bias, classification noise).

Follow-Up Research This Work Enables

1. Aspect-based sentiment analysis to disentangle technology enthusiasm from labor anxiety within the same posts. The paper's word cloud analysis (Figure 13) establishes that positive and negative discourse cluster around different themes, but the forced-choice classification (each post is positive OR negative OR neutral) cannot capture posts that express both technology excitement and unemployment concern. A natural extension is to annotate the same or similar Weibo data with aspect-level labels: for each post, separately label sentiment toward Apollo Go's technology (positive/negative/neutral) and sentiment toward its socioeconomic impact (positive/negative/neutral). This would enable measuring ambivalence — what fraction of posts express mixed attitudes? Does ambivalence increase during the viral surge as a broader audience with more diverse concerns enters the discussion? Are certain provinces (perhaps those with transportation-heavy employment) characterized by higher rates of ambivalence rather than purely negative sentiment? The paper's existing annotation framework and BERT fine-tuning pipeline could be extended to multi-label or aspect-based classification with modest additional annotation effort (perhaps 500–1,000 posts labeled on both dimensions). A strong follow-up would report: (a) the proportion of posts that are technology-positive but labor-negative, (b) whether this proportion changes from January–June to July, and (c) whether provinces with high "negative" sentiment in the current paper are actually characterized by high ambivalence rather than outright opposition — which would carry very different policy implications.

2. Post-July 2024 longitudinal monitoring to distinguish transient viral sentiment shifts from durable attitude changes. The paper's central temporal finding — that negative sentiment proportion increased during the July 2024 viral surge — is based on data truncated at July 14 for the daily analysis (Table 5) and July 31 for the monthly totals (Table 4). We do not know what happened after July. Did the negative sentiment proportion revert to pre-surge levels as attention waned, suggesting the July spike was a temporary influx of non-user discussants whose sentiment differed from the user base? Or did negative sentiment remain elevated in subsequent months, suggesting that the viral moment permanently reshaped public attitudes? A straightforward follow-up would collect Weibo data for August–December 2024 using the same API query and apply the already-trained model to classify new posts. If the positive-to-negative ratio in August returns to ~2.5:1 (matching January–June), the interpretation is that the July negativity spike was a compositional effect of audience expansion rather than genuine attitudinal change. If the ratio stays compressed at ~1.3:1 through December, the interpretation is that the national conversation permanently altered public perception — a much more consequential finding for policymakers. The existing model and code pipeline make this a low-effort, high-value extension.

3. Cross-province regression of sentiment against aggregate demographic and economic indicators to generate hypotheses about geographic drivers. The paper's most striking finding — the Xinjiang-Tibet sentiment divergence — is purely descriptive; the paper cannot explain it because it lacks individual-level demographic data. A strong follow-up would compile province-level aggregate statistics from Chinese statistical yearbooks (e.g., provincial GDP per capita, urbanization rate, median age, employment share in transportation/warehousing, internet penetration rate, average educational attainment) and model the relationship between these provincial characteristics and the sentiment proportions reported in the paper. This is ecological inference — correlating aggregate sentiment with aggregate demographics, which does not establish individual-level relationships — but it can generate testable hypotheses. For example, if provinces with higher transportation-sector employment shares show systematically more negative sentiment, this would support the interpretation that labor market anxiety drives geographic variation. If provinces with higher internet penetration show systematically different sentiment patterns, this would suggest that online information exposure (rather than labor market structure) is the mechanism. The paper's spatial maps (Figures 9–10) provide the dependent variable; publicly available provincial statistics provide the independent variables. A regression with ~30 provincial units would be underpowered for strong causal claims but would provide the first systematic evidence about which provincial characteristics correlate with robotaxi sentiment in China.

4. Multi-platform validation to assess whether Weibo sentiment generalizes to other Chinese social media ecosystems or reflects platform-specific user demographics. All findings are from a single platform. Weibo's user base skews toward younger, urban, and more educated demographics relative to the Chinese population; the sentiment distribution on Weibo may not represent public sentiment on platforms with different user bases (e.g., Douyin for short-video discourse, Zhihu for longer-form analytical discussion, Xiaohongshu for lifestyle-oriented content). A comparative study would collect Apollo Go-related posts from 2–3 additional Chinese platforms over the same time period (January–July 2024), apply platform-specific fine-tuning (since each platform has distinct linguistic norms), and compare the resulting sentiment distributions. If all platforms show the same temporal surge structure (89% of posts in July) and the same Xinjiang-Tibet divergence in spatial sentiment, confidence in the platform-independence of the findings increases. If, say, Zhihu shows more negative sentiment overall (consistent with its more critical/analytical user base) while Xiaohongshu shows more positive sentiment (consistent with its lifestyle-enthusiast orientation), this would reveal that "public sentiment" as measured by any single platform is platform-contingent, and that policymakers should triangulate across platforms rather than relying on any one data source. The paper's code release makes replication on new platforms straightforward — the primary cost is platform-specific annotation effort.

5. User-level longitudinal analysis: do individual users' sentiments change after the viral surge, or does the aggregate shift reflect a changing user composition? The paper treats each post as an independent observation and aggregates sentiment by time period and province. But the data contains User ID fields (Table 1), meaning it is possible to track whether the same users post in both the pre-surge and surge periods, and whether their sentiment shifts. This would directly test the compositional explanation for the negative sentiment increase: if most users who posted in January–June also post in July and their individual sentiment becomes more negative, the shift reflects genuine attitudinal change. If the January–June users largely disappear from the discourse in July (swamped by a new, larger cohort of users with a more negative baseline sentiment), the shift is compositional. A follow-up study would: (a) identify users who posted at least once in January–June and at least once in July, (b) compute the sentiment distribution of their January–June posts vs. their July posts, (c) compare this within-user shift to the aggregate shift. If the within-user sentiment is stable (early adopters remain positive) while the aggregate shifts negative (due to the influx of new users), this would have direct policy implications: the experience-based sentiment of actual users remains positive, suggesting that concerns about public backlash are overstated and that expanding deployment to give more people direct experience might shift the national discourse back toward positivity.

6. Event-correlated sentiment analysis: do specific news events or incidents drive the temporal sentiment patterns? The paper identifies May 15, 2024 (Apollo Day 2024 event in Wuhan) as a potential trigger for the subsequent viral surge, and notes that July 8–12 marks the peak of the surge, but does not attempt to correlate daily sentiment fluctuations with specific external events. Apollo Go likely experienced news events during this period — new city launches, pricing changes, safety incidents, regulatory announcements, competitor actions (from Pony.ai, AutoX, WeRide, or Waymo/Tesla as international comparators), or statements by public officials — that could drive the observed sentiment dynamics. A follow-up study would compile a timeline of Apollo Go-related news events from January–July 2024 (from news archives, Baidu corporate announcements, and transportation regulatory filings) and conduct an interrupted time-series analysis: does sentiment (positive proportion, negative proportion) change significantly in the 24–48 hours following specific event types? If negative sentiment spikes are tightly coupled to specific incident types (e.g., an accident, a viral video of a robotaxi blocking traffic, a labor union statement), the policy response should be incident-specific (improving the specific operational domain that failed) rather than general (slowing overall deployment). If sentiment shifts are diffuse and not event-correlated, the driver is likely broader media narrative dynamics rather than specific operational failures.

Practical Applications and Downstream Use Cases

1. City-level regulatory decisions about robotaxi deployment permits. Cities currently evaluating whether to allow Apollo Go or competing robotaxi services to operate commercially must assess public acceptance — a political prerequisite for permitting deployment. The standard approach is to conduct surveys, which are slow (months from design to results), expensive, and typically sample a few hundred residents. The methodology demonstrated in this paper — fine-tuning a BERT model on ~2,700 labeled social media posts and applying it to tens of thousands of unlabeled posts — can produce a geographically resolved, temporally granular sentiment map in weeks rather than months. A city transportation department in, say, Chengdu (which does not yet have Apollo Go according to Table 3, but is in Sichuan, which ranked 9th in July discussion volume with 798 posts) could use this pipeline to monitor: (a) whether discussion volume is rising (indicating growing public awareness), (b) what the baseline sentiment in Sichuan is before deployment (informing outreach strategy), and (c) whether sentiment shifts after deployment begins (providing real-time feedback on public response). The specific benefit: the paper's finding that negative sentiment increased from ~22% to ~33% between pre-surge and surge periods means that a city monitoring sentiment during a viral moment would see a worse picture than during quiet periods — understanding this temporal dynamic prevents overreaction to transient negativity spikes.

2. Targeted public communication strategy for service providers. The word cloud analysis (Figure 13) demonstrates that negative sentiment clusters around unemployment concerns rather than safety fears, at least in the Chinese Weibo discourse analyzed. For Baidu (Apollo Go's operator), this finding has direct implications for public communication strategy. If safety were the dominant concern, Baidu should invest in publishing safety statistics, offering public ride-alongs, and emphasizing sensor redundancy and remote operator oversight. But if unemployment is the dominant concern, safety messaging addresses the wrong anxiety. Instead, Baidu should communicate about workforce transition programs (retraining for traditional taxi drivers), new job categories created by autonomous vehicle operations (fleet maintenance, remote monitoring, customer support), and the economic benefits of lower-cost transportation for workers in other sectors. The paper's finding that some positive posts express frustration with traditional taxi driver behavior (smoking, complaining about short trips) also suggests a messaging angle: frame Apollo Go not as replacing human drivers but as providing an alternative for riders who are underserved or mistreated by the incumbent service — a framing that acknowledges the labor concern while redirecting the comparison. The specific benefit: data-driven message targeting that addresses the actual concerns expressed in spontaneous public discourse rather than the concerns that engineers or executives assume the public holds.

3. Early-warning system for transportation agencies monitoring social equity impacts of autonomous mobility. One of the paper's findings — that provinces with high discussion intensity correlate with operational provinces but that sentiment varies dramatically even among non-operational provinces — provides a template for an equity monitoring framework. A transportation agency could deploy this pipeline on an ongoing basis to track which geographic regions and which thematic concerns are dominating the public conversation, and whether marginalized or rural regions (which may not receive early robotaxi deployment) are expressing concerns that need policy attention before deployment reaches them. For example, the paper's finding that Gansu and Tibet — western, less economically developed provinces — show predominantly negative sentiment toward Apollo Go raises an equity question: are these provinces concerned that autonomous ride-hailing will widen the transportation access gap between wealthy coastal cities (which get the service first) and poorer interior regions (which may wait years for deployment)? If so, the policy response might be to include western provinces in early deployment planning or to ensure that traditional transit services are not degraded in anticipation of autonomous services that may arrive slowly. The specific benefit: moving from reactive sentiment monitoring ("what are people saying now?") to proactive equity analysis ("which populations are expressing concerns, and do those concerns align with structural transportation inequities that policy should address?").

4. Benchmarking tool for cross-national comparative studies of autonomous vehicle public acceptance. The paper notes repeatedly that its findings differ from Western studies in specific ways — privacy and data security concerns are prominent in U.S. and European AV discourse (Chen and Tomblin, 2021) but absent from Chinese Weibo posts about Apollo Go, while unemployment anxiety dominates Chinese negative sentiment but is one of several concerns in Western contexts. This observation remains anecdotal in the current paper because no formal cross-platform comparison is conducted. However, the paper's released model and annotation protocol provide a standardized instrument: future researchers studying AV sentiment in other countries could apply the same fine-tuning approach (Chinese BERT for Chinese platforms; multilingual BERT or language-specific models for other languages) and the same sentiment category definitions, enabling direct comparison. If a study of Waymo discourse on Twitter using a comparable three-way classification (positive/negative/neutral) found that 45% of negative posts concerned safety while only 15% concerned employment — the inverse of the Apollo Go pattern — this would establish that the structure of public concern varies systematically by national context, which has direct implications for how global AV companies should tailor their public engagement by market. The specific benefit: a replicable measurement instrument that enables the accumulation of comparable evidence across countries, platforms, and services, transforming a series of isolated case studies into a cumulative research program on the social acceptance of autonomous mobility.

When to Prefer This Method

The paper does not position its methodology against explicit named alternatives (e.g., "use fine-tuned BERT rather than lexicon-based sentiment scoring because X" or "use Weibo data rather than surveys because Y"). It presents its pipeline as a demonstration that NLP tools can extract transportation-policy-relevant insights from social media data, not as a comparative claim that this approach dominates alternatives. The paper acknowledges several limitations — lack of demographic data, short time span, single-platform scope — that implicitly suggest situations where traditional survey methods would be preferable (e.g., when you need to link sentiment to individual demographic characteristics or when you need representative population sampling rather than social media self-selection). However, the paper does not articulate a decision framework or explicit tradeoff between its approach and alternatives, so a structured "prefer A when, prefer B when" matrix would be fabricating a comparison the paper itself does not make. Practitioners should understand this work as providing a complementary measurement tool — fast, geographically granular, responsive to real-time events — that augments rather than replaces survey-based public acceptance research, with the choice between methods depending on whether the research question prioritizes representativeness and demographic resolution (surveys) or scale, geographic granularity, and temporal responsiveness (social media NLP).