SentiWeibo Sentiment Analysis

• SentiWeibo is a Weibo-specific sentiment analysis framework that couples micro-level personalized responses with macro-level emotional distribution modeling.
• It employs a dual-stage data curation process, combining hashtag-based topic selection with extensive user history crawling and human filtering.
• The framework spans diverse labeling schemes—from binary polarity to nuanced seven-emotion classifications—informing applications from crisis tracking to market forecasting.

SentiWeibo denotes Weibo-specific sentiment analysis in two closely related senses. In the narrow sense, it is the large-scale dataset introduced for dual-level public response prediction on authentic Sina Weibo interactions, with 53 topics, 7,837 hashtagged posts from 6,401 unique users, 476,605 historical posts, and official topic-level sentiment distributions over seven emotions. In the broader research literature, the term also refers to the task family of sentiment, stance, emotion, and public-response modeling on Weibo, spanning binary stance detection, ternary sentiment classification, four-way sentiment analysis with sarcasm, personalized response generation, and network-level emotion dynamics (Zhang et al., 1 Aug 2025).

1. Terminological scope and historical development

The named SentiWeibo dataset emerged in work on dual-level public response prediction, where the central objective is to connect individual-level generated responses with macro-level public sentiment distributions. That formulation makes SentiWeibo not merely a corpus of labeled posts, but a benchmark for coupling personalization, generation, and aggregate sentiment alignment. The same work defines the dataset around hashtagged topics, user histories, derived personas, and official Weibo sentiment distributions, thereby moving beyond conventional post-level polarity classification (Zhang et al., 1 Aug 2025).

Earlier Weibo sentiment studies, however, establish a broader lineage in which “SentiWeibo” is best understood as a research setting rather than a fixed canonical resource. Early work analyzed four emotions—anger, joy, sadness, and disgust—on approximately 70 million Weibo posts and an interaction network of 9,868 nodes and 19,517 edges at threshold $T=30$, emphasizing sentiment correlation and propagation rather than only classification (Fan et al., 2013). Topic-specific stance analysis on Traditional Chinese Medicine used user tags for distant supervision and trained a support vector machine over TCM-related tweets, treating Weibo as a domain-specific opinion source rather than a generic sentiment benchmark (Shen et al., 2014). Later studies operationalized Weibo sentiment analysis as binary classification over legal-event comments with LSTM models, three-class sentiment classification with CNNs on “weibo senti 100k,” and four-class crisis sentiment analysis with sarcasm as a separate label (Wang et al., 2022, Xie et al., 2023, Hu et al., 9 Jan 2025).

This trajectory shows an expansion from post-level polarity detection to richer problem formulations: stance, discrete emotion taxonomies, sarcasm-aware classification, macro sentiment distributions, and sentiment evolution under social influence. A plausible implication is that SentiWeibo is now less a single benchmark name than a layered research program centered on Sina Weibo as a source of public opinion signals.

2. Data construction, curation, and annotation regimes

The named SentiWeibo dataset is built from Sina Weibo hashtag discussions obtained from two channels: Weibo Hot Topics and Weibo AI Search. The authors curated 53 hashtags/topics, collected hundreds of posts for each hashtag, and paired them with the corresponding public sentiment distributions provided by Weibo. For each participating user, they crawled historical posts through a web spider built with Scrapy and Selenium. Cleaning is two-stage: human annotators remove off-topic hashtag posts and posts from mainstream media accounts, while user histories are filtered by retaining hashtagged posts and removing posts with noise-indicating keywords. Human evaluation on 100 random cases found only 1% of the retained history posts irrelevant or noisy. The final resource contains 7,837 hashtagged posts, 6,401 unique users, and 476,605 historical posts, with an 8/1/1 train/validation/test split and 10 to 953 history posts per user (Zhang et al., 1 Aug 2025).

Other Weibo sentiment resources use markedly different supervision strategies. The CNN study on “weibo senti 100k” begins from 119,988 Simplified Chinese Weibo posts from Baidu’s PaddlePaddle AI platform, originally labeled as 59,993 positive and 59,995 negative. After cleaning and removing empty posts, 117,282 posts remain. The authors then rescore each cleaned post with SnowNLP and remap sentiment scores $s \in [0,1]$ into a three-class scheme: negative for $s < 0.3$, neutral for $0.3 \le s \le 0.7$, and positive for $s > 0.7$. The resulting class distribution is strongly imbalanced: 72,909 positive, 26,368 negative, and 18,005 neutral (Xie et al., 2023).

Weak supervision by external metadata also appears in the TCM stance study. That pipeline first collects 48,861 Weibo users through profile tags such as “中医,” retrieves 21,242,370 tweets through official Sina Weibo APIs, expands retweets to reach 43,012,068 tweet units, and then filters to 1,650,497 TCM-topic tweets by requiring at least two distinct TCM terms. User stance is inferred from self-assigned tags such as “爱中医” or “反中医,” yielding 1,866 supporting-TCM users and 290 opposing-TCM users, whose 40,888 and 6,975 TCM tweets respectively become the labeled training set (Shen et al., 2014).

Event-centric studies adopt still other constructions. The LSTM case study crawls Sina Weibo comments concerning the Tangshan attack and the Jiang Ge case, labels them automatically with SnowNLP, and uses Tangshan-related reviews for training and validation and Jiang Ge-related reviews for testing (Wang et al., 2022). The COVID-19 study starts from the Harvard Dataverse Weibo COVID dataset, with $N_{\text{total}} = 4{,}049{,}407$ posts, distinguishes distinct posts, duplicate posts, and reposts, and then labels the corpus with Llama 3 8B after validating on a 199-post gold set produced by OpenAI o1-mini with native-speaker verification (Hu et al., 9 Jan 2025).

Taken together, these pipelines define a characteristic SentiWeibo pattern: large-scale Weibo collection, aggressive preprocessing for microblog noise, and supervision through a mix of platform signals, weak labels, distant supervision, or manual filtering rather than purely manual annotation.

3. Task formulations and label spaces

Weibo sentiment analysis under the SentiWeibo umbrella is notable for the diversity of its label spaces. The research record includes binary stance, binary sentiment, ternary sentiment, four-way sentiment with sarcasm, four-emotion network analysis, and seven-emotion macro distribution modeling.

Work	Label space	Task
Named SentiWeibo	happy, sad, angry, calm, fear, surprised, disgusted	Micro-level response prediction; macro-level sentiment distribution prediction
CNN on “weibo senti 100k”	negative, neutral, positive	Three-class sentiment classification
COVID-19 Weibo study	positive, negative, sarcastic, neutral	Non-binary crisis sentiment classification
LSTM legal-event study	positive, negative	Event-centric public-opinion classification
TCM stance study	supporting TCM, opposing TCM	Topic-specific stance detection
Emotion-correlation study	anger, joy, sadness, disgust	User-level emotion correlation in interaction networks

For the named SentiWeibo dataset, the micro-level unit is a user-conditioned response to a topic, while the macro-level unit is a topic-level sentiment distribution. Formally, each user $u$ has a history

$$\mathcal{H}_u = \{(x_u^1, y_u^1), (x_u^2, y_u^2), \cdots, (x_u^{N_u}, y_u^{N_u})\},$$

and the goal is to generate a personalized response $y_u^{\prime\,\mathcal{N}}$ for a topic $x^{\mathcal{N}}$, then aggregate generated sentiments into a predicted distribution $s \in [0,1]$0 to compare against Weibo’s official distribution $s \in [0,1]$1. The official taxonomy is

$s \in [0,1]$2

This formulation departs sharply from standard sentence classification because the output space includes both text generation and distribution alignment (Zhang et al., 1 Aug 2025).

By contrast, the CNN study turns a balanced binary resource into a ternary task through SnowNLP thresholds, thereby inserting a neutral class into “weibo senti 100k.” The COVID-19 study explicitly isolates sarcasm as an independent category rather than subsuming it under negative polarity, arguing that exaggerated praise and pretend naivete are essential to Chinese crisis discourse on Weibo. The TCM study is best described as stance detection, because the classes concern support or opposition to Traditional Chinese Medicine rather than generic positive or negative sentiment. The 2013 network study is an emotion-correlation analysis rather than a conventional classifier, since each user is represented by an emotion vector $s \in [0,1]$3 built from fractions of tweets in four emotions (Xie et al., 2023, Hu et al., 9 Jan 2025, Shen et al., 2014, Fan et al., 2013).

4. Modeling architectures and methodological paradigms

SentiWeibo research spans classical linear models, sequence models, CNN text classifiers, transformer-based classifiers, prompted LLMs, personalized LLMs, and opinion-dynamics systems.

The TCM stance system is architecturally simple but methodologically distinctive. Tweets are converted from Traditional Chinese to Simplified Chinese, cleaned of URLs, mentions, emoticons, and advertisements, segmented with ICTCLAS using a 5,307-word custom dictionary, and represented as binary unigram features. Chi-square feature selection retains 3,000 features, and a linear SVM is trained with class-weight tuning through parameter $s \in [0,1]$4. A post-classification user-consistency adjustment uses

$s \in [0,1]$5

where $s \in [0,1]$6 and $s \in [0,1]$7 are counts of predicted supporting and opposing tweets for a user (Shen et al., 2014).

The troll-detection pipeline is more hybrid. It implements HMM-based Chinese segmentation with B/M/E/S tags, Word2Vec embeddings, Naïve Bayes-derived word sentiment scores, six emotion scores based on MI, CHI, and TF-IDF features, and user/activity features such as follower count, following count, urank, verified, floor_number, like_count, and sentiment-difference features. These are then consumed by XGBoost or SVM classifiers, with comment-level sentiment among the most important features (Jiang et al., 2021).

Neural sequence modeling appears in the legal-event LSTM study, where character-level segmentation is used because Chinese has no spaces and the data size is limited. Word2Vec embeddings of dimension 300 feed an LSTM with dropout, Adam, and early stopping, followed by a sigmoid classifier for binary sentiment (Wang et al., 2022). The CNN study on “weibo senti 100k” uses Keras tokenization with num_words = 5000, sequence padding with maxlen = 400, a trainable embedding layer with $s \in [0,1]$8, $s \in [0,1]$9, $s < 0.3$0, a Conv1D layer with 250 filters of width 3 and ReLU activation, GlobalMaxPooling1D, a dense layer of size 250, dropout 0.2 at two points, and a 3-way softmax trained with categorical_crossentropy and Adam (Xie et al., 2023).

Transformer-based sentiment engines appear in two distinct ways. In financial forecasting, Chinese BERT from Google’s official BERT release is fine-tuned on weibo_senti_100k with batch size 64, learning rate $s < 0.3$1, 3 epochs, and warm-up proportion 0.1, then combined with a two-layer LSTM of hidden size 128 for Hang Seng Index prediction (Ma et al., 2024). In the opinion-dynamics work on influential tech bloggers, bert-base-chinese fine-tuned on weibo_senti_100k is selected as the sentiment extractor because it achieves AUC $s < 0.3$2, F1 $s < 0.3$3, and Accuracy $s < 0.3$4 on a separate 2018 Weibo test set, after which comment-level scores $s < 0.3$5 are aggregated to post-level and blogger-level sentiment trajectories (He et al., 19 Nov 2025).

The LLM-based COVID-19 study eliminates task-specific training altogether at inference time. Llama 3 8B Instruction-tuned is used as a few-shot classifier over raw Chinese text, with prompt instructions that require output from the fixed set {positive, negative, sarcastic, neutral} and explicitly distinguish [reply]//[original post] structure by directing the model to classify the reply content only (Hu et al., 9 Jan 2025).

The named SentiWeibo benchmark introduces the most elaborate architecture. SocialAlign uses SocialLLM with a Personalized Analyze-Compose LoRA structure, retrieves topic-relevant user history with BM25, derives a five-dimensional persona with GPT-4, and generates user-specific responses according to

$s < 0.3$6

At the macro level, predicted sentiments are aggregated into a distribution $s < 0.3$7, and the optimization target is alignment with the official distribution $s < 0.3$8 through Jensen–Shannon divergence (Zhang et al., 1 Aug 2025).

Finally, SentiWeibo research extends beyond classification into dynamical systems. The sentiment-evolution study models each blogger’s audience as a macro-agent with expressed opinion $s < 0.3$9 and, in two-layer variants, private opinion $0.3 \le s \le 0.7$0. It evaluates French–DeGroot, Friedkin–Johnsen, FDG with memory, and expressed–private opinion models, including delayed variants, over a six-month Huawei-focused Weibo panel (He et al., 19 Nov 2025).

5. Empirical performance and observed benchmark behavior

On the named SentiWeibo benchmark, SocialAlign and PAC-LoRA outperform prompted LLMs, LoRA fine-tuning, PER-PCS, and other baselines on both micro-level and macro-level criteria. On the three demonstration topics, PAC-LoRA reaches Sentiment Acc 35.6, F1 38.5, and JS divergence 34.1 for Public Health; Acc 43.5, F1 50.2, and JS 26.1 for Recruitment Policies; and Acc 48.2, F1 50.1, and JS 17.5 for Financial Scams. Human evaluation also favors PAC-LoRA on Language Style, Content Focus, and Persona Dynamics (Zhang et al., 1 Aug 2025).

Earlier SentiWeibo-style classifiers report strong but task-dependent performance. The CNN on “weibo senti 100k” attains macro-average F1 $0.3 \le s \le 0.7$1, with per-class values of 0.76 for negative, 0.65 for neutral, and 0.77 for positive; neutral is the hardest class. The binary LSTM case study reports Loss 0.2627, MAE 0.0916, Accuracy 92.06%, Precision 92.08%, and Recall 92.04% on Jiang Ge case reviews. The COVID-19 Llama 3 8B classifier reports weighted F1 $0.3 \le s \le 0.7$2, with per-class F1 of 0.4286 for sarcastic, 0.7977 for neutral, 0.7692 for negative, and 0.8082 for positive, confirming that sarcasm is the most difficult category. In specialized domains, the TCM stance study reports F-measure 97%, the troll-detection pipeline reaches 89% accuracy with XGBoost under one configuration, and the financial sentiment pipeline reaches 87% accuracy for the AFA group versus 62% for the UFA group, with AFA precision 39.67% higher in relative terms (Xie et al., 2023, Wang et al., 2022, Hu et al., 9 Jan 2025, Shen et al., 2014, Jiang et al., 2021, Ma et al., 2024).

A second line of results concerns not classification accuracy but social behavior. The 2013 emotion-correlation study finds that anger is significantly more correlated among users than joy, while sadness is low and highly fluctuated; stronger interactions and larger degree are associated with stronger sentiment correlation, especially for anger (Fan et al., 2013). The sentiment-evolution study over Huawei-focused blogs finds that raw Weibo sentiment trajectories violate instantaneous averaging but become much more consistent with delayed averaging when $0.3 \le s \le 0.7$3 or $0.3 \le s \le 0.7$4. Among the tested dynamical models, reduced EPO and delayed EPO variants yield the best overall fit or prediction under different criteria, whereas plain French–DeGroot performs worst (He et al., 19 Nov 2025).

These results indicate that SentiWeibo is empirically heterogeneous. High scores are achievable in tightly defined binary or topic-specific tasks, whereas ternary, sarcasm-aware, or seven-emotion personalized benchmarks remain materially harder.

6. Applications, limitations, and future directions

SentiWeibo systems are used for social media monitoring, market research, brand management, public-opinion supervision, policy analysis, crisis sentiment tracking, stock-market forecasting, troll detection, and modeling of influencer-audience sentiment evolution. The CNN study explicitly highlights social media analysis, market research, and policy studies; the legal-event LSTM frames the task as network public opinion analysis; the stock-market study uses Weibo sentiment to predict the Hong Kong Hang Seng index; the troll-detection work targets real-time identification of Water Army comments in a Chrome extension; and the opinion-dynamics study focuses on influencers, marketers, and follower communities (Xie et al., 2023, Wang et al., 2022, Ma et al., 2024, Jiang et al., 2021, He et al., 19 Nov 2025).

Several limitations recur across this literature. Weak supervision is pervasive: SnowNLP rescoring, user-tag-based stance inference, LLM-assisted labeling, and platform-provided macro sentiment distributions each introduce their own biases. Class imbalance is common, as in the three-class CNN resource where positive posts dominate after relabeling. Subtle linguistic phenomena remain difficult: the CNN study identifies neutral as the hardest class, while the COVID-19 study reports the lowest F1 for sarcasm and ties this to exaggerated praise, coded criticism, and moderation-aware rhetoric. Dataset construction is also topic-biased or activity-biased: the named SentiWeibo benchmark emphasizes trending and emerging hashtags, active participants, and personal perspectives after removal of mainstream media posts, while some event-centric corpora are not publicly released. In financial forecasting, user expertise is inferred heuristically from verification metadata and finance-related keywords, and no formal significance tests are reported. In dynamical modeling, linear time-invariant opinion models abstract away exogenous shocks and intra-community heterogeneity (Zhang et al., 1 Aug 2025, Hu et al., 9 Jan 2025, Ma et al., 2024, He et al., 19 Nov 2025).

The stated future directions are correspondingly broad. The CNN study proposes RNNs, transformers, and more complex pre-trained models such as BERT; the LSTM case study suggests expansion beyond binary, domain-specific setups; the COVID-19 work motivates stronger sarcasm handling and richer non-binary schemes; the named SentiWeibo benchmark opens a path toward stronger persona extraction, better micro-to-macro alignment, and broader computational social science questions; and the sentiment-evolution work points toward heterogeneous agents, multi-topic modeling, nonlinear dynamics, exogenous event variables, and richer NLP such as aspect-based sentiment and sarcasm detection (Xie et al., 2023, Wang et al., 2022, Hu et al., 9 Jan 2025, Zhang et al., 1 Aug 2025, He et al., 19 Nov 2025).

SentiWeibo therefore occupies a distinctive place in Chinese NLP and computational social science. As a named dataset, it formalizes dual-level public response prediction with authentic social interactions and official sentiment distributions. As a broader research tradition, it encompasses the full spectrum of Weibo sentiment analysis, from stance-labeled topic corpora and resource-efficient CNN baselines to LLM-based sarcasm-aware labeling, persona-conditioned response generation, and networked models of emotional contagion and sentiment evolution.