QNQCDE: Efficient Dialogue Embeddings Based on Contrastive Learning Using Question and Non-Question Pairs

초록

This paper presents QNQCDE, a contrastive learning approach for dialogue embedding that generalizes across both multi-turn and multi-party dialogues. Unlike previous methods relying on PLMs or limited to two-party settings, QNQCDE models interactions between Questions and Non-Questions (QNQ) to capture essential dialogue structures. By defining three types of Q-NQ relatedness, it generates positive, soft-positive, and negative pairs for fine-grained representation learning. Experiments on six datasets across domain categorization, semantic relatedness, and dialogue retrieval show state-of-the-art performance, improving purity, Spearman's correlation, and MAP by 6.1%, 3.0%, and 4.5%, respectively. Code and data are available at this URL. © 2025 IEEE.

키워드