PVGRU: Generating Diverse and Relevant Dialogue Responses via Pseudo-Variational Mechanism

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3295–3310 July 9-14, 2023 ©2023 Association for Computational Linguistics PVGRU: Generating Diverse and Relevant Dialogue Responses via Pseudo-Variational Mechanism Yongkang Liu1,2,3, Shi Feng1, Daling Wang1, Yifei Zhang1, Hinrich Schütze2,3 1 Northeastern University, China 2 Center for Information and Language Processing, LMU Munich 3 Munich Center for Machine Learning (MCML), LMU Munich misonsky@163.com, {fengshi,wangdaling,zhangyifei}@cse.neu.edu.cn Abstract We investigate response generation for multi- turn dialogue in generative chatbots. Exist- ing generative models based on RNNs (Recur- rent Neural Networks) usually employ the last hidden state to summarize the history, which makes models unable to capture the subtle vari- ability observed in different dialogues and can- not distinguish the differences between dia- logues that are similar in composition. In this paper, we propose Pseudo-Variational Gated Recurrent Unit (PVGRU). The key novelty of PVGRU is a recurrent summarizing vari- able that aggregates the accumulated distri- bution variations of subsequences. We train PVGRU without relying on posterior knowl- edge, thus avoiding the training-inference in- consistency problem. PVGRU can perceive subtle semantic variability through summariz- ing variables that are optimized by two ob- jectives we employ for training: distribution consistency and reconstruction. In addition, we build a Pseudo-Variational Hierarchical Dialogue (PVHD) model based on PVGRU. Experimental results demonstrate that PVGRU can broadly improve the diversity and relevance of responses on two benchmark datasets. 1 Introduction The structure of natural language discourse is com- plex and highly variable (Gormley and Tong, 2015; Chung et al., 2015; Nie et al., 2022); this is es- pecially true for dialogue. As shown in Figure 1, examples (a) and (b) have the same dialogue history but they end with different responses: utterances ua 6 vs. ub 6. On the other hand, two dialogues with semantically similar utterances may express quite different context meanings. Because of this vari- ability, there is no simple one-to-one mapping be- tween dialogue context and response. The mapping can be one-to-many – as in Figure 1, i.e., different responses to the same dialogue context – as well as many-to-one, i.e., different context histories re- quiring the same response. We observe that the distribution of a dialogue context (e.g., N a 6 and N b in the figure) is composed of the distribution of its6 utterances and the distribution of each utterance is composed of the distribution of its words. A good model of word level and utterance level variation is a key requirement for improving the quality of responses in dialogue. One line of research (Henderson et al., 2014; Shang et al., 2015; Serban et al., 2016; Luo et al., 2018) employs recurrent neural networks (RNNs) to model dialogue context. However, standard RNNs are not well suited for dialogue context vari- ability (Chung et al., 2015). This is because the in- ternal transition structure of RNNs is deterministic. Thus, RNNs cannot effectively model randomness and variability in dialogue context (Chung et al., 2015). Variational mechanism has been shown to be well …

Content truncated to 500 words