Human-Machine Multi-Turn Language Dialogue Interaction Based on Deep Learning.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Multi-turn dialogue systems face increasing challenges in intention recognition and response generation as conversations lengthen.
• Effective context information extraction is crucial for improving the performance of sequence-to-sequence (Seq2Seq) encoders in dialogue modeling.

Purpose of the Study:

• To optimize the context information extraction capabilities of Seq2Seq encoders for multi-turn dialogue.
• To propose novel models that better capture historical and current dialogue information for enhanced contextual understanding.

Main Methods:

• Developed a BERT-based fusion encoder, ProBERT-To-GUR (PBTG), integrating historical and current dialogue information.
• Introduced an enhanced ELMO model, 3-ELMO-Attention-GRU (3EAG), with attention mechanisms for improved context extraction.
• Evaluated models on the LCCC-large and Naturalconv multi-turn dialogue datasets.

Main Results:

• Both PBTG and 3EAG models demonstrated significant improvements over state-of-the-art models in open-domain and fixed-topic multi-turn dialogue experiments.
• The 3EAG model achieved an optimal average BLEU score of 32.4 for fixed-topic dialogue, outperforming other models.
• The PBTG model achieved an optimal average BLEU score of 31.8 for open-domain dialogue, showing strong performance.

Conclusions:

• The 3EAG model is particularly effective for fixed-topic multi-turn dialogues, achieving superior language generation.
• The PBTG model demonstrates greater strength in open-domain multi-turn dialogue tasks.
• The proposed models significantly advance the field of multi-turn dialogue research by enhancing contextual understanding and generation capabilities.