Multi-turn Conversation Intent Recognition Incorporating Historical Information
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This paper first constructs a synergistic mechanism that integrates semantic reasoning and local text training. The semantic reasoning model uses dynamic feature transfer and gating control to achieve on-demand filtering and orderly transfer of historical information. Local text training employs sliding window sampling and contrastive learning to enhance the model's ability to capture key contextual semantics. Secondly, conversation-specific pre-training tasks are designed, including conversation coherence prediction (determining the temporal rationality of fragments) and intent mask recovery (focusing on learning intent keywords), to optimize the model's adaptability to multi-turn conversation scenarios. Experiments are conducted on the DSTC7 and MultiWOZ 2.1 datasets. Results show that on the DSTC7 dataset, the model achieves intent recognition accuracy of 92.3% and F1-score of 91.5%, improvements of 4.8 and 5.1 percentage points, respectively, over BERT-Base. On the MultiWOZ 2.1 dataset, the model achieves accuracy of 90.8% and F1-score of 91.7%, improvements of 3.6 and 5.2 percentage points, respectively, over RoBERTa-Base. Research demonstrates that the proposed method can effectively integrate historical information, significantly improving the accuracy and robustness of multi-turn conversation intent recognition, providing technical support for performance optimization of intelligent human-computer interaction systems.
Fields of Study & Subjects
Speech and dialogue systems
Topic Modeling
AI in Service Interactions
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 953–958
- Publication date
- 28 January 2026
- Language
- EN