weibo-opinion-dynamic-sentiment-bert

BERT Sentiment Model for Chinese Weibo Text

This repository contains a TensorFlow BERT model fine-tuned for binary sentiment classification of Chinese Weibo-style short text.

The model is associated with the paper:

Opinion Dynamics Models for Sentiment Evolution in Weibo Blogs
Yulong He, Anton V. Proskurnikov, Artem Sedakov
arXiv:2511.15303
https://arxiv.org/abs/2511.15303

The model is used to estimate sentiment scores from Weibo comments. The positive-class probability can be interpreted as a continuous sentiment score in [0, 1].

Model Description

• Model type: BERT sequence classification model
• Framework: TensorFlow
• Base model: bert-base-chinese
• Task: Binary sentiment classification
• Language: Chinese
• Domain: Weibo-style short text and Chinese social media comments
• Labels:
  • 0: negative
  • 1: positive
• Output: Class probabilities for negative and positive sentiment

The model was fine-tuned from bert-base-chinese using the Hugging Face Transformers library.

Intended Use

This model is intended for:

• Chinese sentiment analysis
• Weibo comment sentiment classification
• Social media text classification
• Estimating positive sentiment probability for Chinese short text
• Research on sentiment evolution and opinion dynamics

Example research applications include:

• assigning sentiment scores to individual Weibo comments
• aggregating comment-level sentiment into post-level sentiment
• analyzing collective sentiment evolution over time
• reproducing the sentiment extraction step in the associated opinion-dynamics paper

Out-of-Scope Use

This model is not intended for:

• high-stakes decision making
• psychological diagnosis
• individual-level profiling
• moderation without human review
• non-Chinese text classification
• fine-grained emotion classification beyond binary positive/negative sentiment

Training Data

The model is based on bert-base-chinese and fine-tuned on Chinese sentiment annotation data.

This model fine-tuned on: weibo_senti_100k
A Weibo sentiment dataset containing more than 100,000 anonymous Chinese Weibo texts with positive and negative sentiment labels.

Citation

If you use this dataset, please cite:

@article{He2026Mapping,
  author  = {He, Yulong and Proskurnikov, Anton V. and Sedakov, Artem},
  title   = {Mapping sentiment dependencies among online influencer communities on {Weibo}},
  journal = {Social Network Analysis and Mining},
  year    = {2026},
  month   = sep,
  doi     = {10.1007/s13278-026-01645-w},
  url     = {https://doi.org/10.1007/s13278-026-01645-w},
  issn    = {1869-5469}
}