CobSeg: New AI Model for Dialogue Topic Segmentation
A novel AI model named CobSeg has been introduced for the task of dialogue topic segmentation, which is essential for human-AI collaboration. This model tackles the difficulty of recognizing various boundary cues, such as lexical transitions at the edges of utterances and semantic breaks between them. Traditional utterance models often overlook these localized lexical signals. CobSeg features an innovative multi-branch design that differentiates between semantic continuity and lexical transitions, recovering both through directional boundary predictions. Additionally, it applies boundary informativeness weighting to highlight significant utterance positions and integrates a topic coherence cue derived from a corpus with learned weights. Evaluated on five benchmark datasets, CobSeg shows superior boundary prediction without requiring LLM calls during inference. The findings are documented on arXiv with the identifier 2605.30668.
Key facts
- CobSeg is a multi-branch architecture for dialogue topic segmentation.
- It separates coherence-level semantic continuity from lexical boundary transitions.
- Uses directional boundary prediction and boundary informativeness weighting.
- Incorporates corpus-derived topic coherence cue with learned weights.
- Evaluated under supervised gold-boundary training and pseudo-label setting.
- Performs enhanced boundary prediction without LLM calls during inference.
- Tested across five benchmark datasets.
- Published on arXiv with identifier 2605.30668.
Entities
Institutions
- arXiv