One-to-Many and Many-to-One Dialogue Learning via Sentence Semantic Segmentation Guided Conditional Variational Auto-Encoder

Anonymous

Abstract: Due to the complex mapping relations, one-to-many and many-to-one phenomena are huge challenges for open-domain dialogue generation task, which tend to make dialogue models generate irrelevant, incoherent or non-diverse responses. Most existing methods avoid learning such phenomena through introducing the external information, reconstructing the optimization function or manipulating data samples. However, avoiding confronting such challenges ignores valuable information in these responses, and the dialogue models cannot learn the nature of such phenomena. In this paper, we propose a Sentence Semantic Segmentation guided Conditional Variational Auto-Encoder (SegCVAE) to directly learn one-to-many and many-to-one responses. SegCVAE uses prominent semantics to replace the original semantics to learn the distribution of latent variables, which avoids the gap between latent variables and the context, thus ensuring the relevance and coherence of the generated responses. Furthermore, SegCVAE can segment multiple prominent semantics to ensure the diversity of generated responses. To evaluate the model, we first define two new tasks named one-to-many dialogue learning task and many-to-one dialogue learning task. And then provide two new dialogue datasets named One-to-Many and Many-to-One, which are extracted from the well-established dataset. Finally, we also propose the evaluation strategies based on some commonly-used metrics. The experiment results show that our model achieve better performance than the baseline models in addressing these two new tasks.