Reproduce “Topic Shift Detection in Chinese Dialogues: Corpus and Benchmark” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiangyi Lin, Yaxin Fan, Feng Jiang, Xiaomin Chu, Peifeng Li
A new Chinese Natural Topic Dialogue (CNTD) corpus and a benchmark are introduced for dialogue topic shift detection. The proposed teacher-student framework with hierarchical contrastive learning achieved an F1-score of 53.9% on CNTD, improving 5.0% over the best baseline, and demonstrated generalizability with a 5.8% F1 improvement on an English dataset.
Key insights
- Implicitly simulating future response information through techniques like Information Simulation Loss can significantly improve topic shift detection in real-time, response-unknown scenarios.
- Hierarchical Contrastive Learning effectively enhances model performance by promoting semantic coherence within topic categories and helping to mitigate the challenges of data imbalance.
- Developing a large-scale, naturally annotated dialogue corpus is fundamental for training and evaluating robust models for complex conversational AI tasks like topic shift detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.