Diverse dialogue generation with context dependent dynamic loss function

Proceedings of the 28th International Conference on Computational Linguistics, pages 4123–4127 Barcelona, Spain (Online), December 8-13, 2020 4123 Diverse dialogue generation with context dependent dynamic loss function Ayaka Ueyama, Yoshinobu Kano Faculty of Informatics, Shizuoka University, Japan aueyama@kanolab.net, kano@inf.shizuoka.ac.jp Abstract Dialogue systems using deep learning have achieved generation of fluent response sentences to user utterances. Nevertheless, they tend to produce responses that are not diverse and which are less context-dependent. To address these shortcomings, we propose a new loss function, an In- verse N-gram Frequency (INF) loss, which incorporates contextual fluency and diversity at the same time by a simple formula. Our INF loss can adjust its loss dynamically by a weight using the inverse frequency of the tokens’ n-gram applied to Softmax Cross-Entropy loss, so that rare tokens appear more likely while retaining the fluency of the generated sentences. We trained Transformer using English and Japanese Twitter replies as single-turn dialogues using different loss functions. Our INF loss model outperformed the baselines of SCE loss and ITF loss models in automatic evaluations such as DIST-N and ROUGE, and also achieved higher scores on our human evaluations of coherence and richness. 1 Introduction Recently, many reports have described studies using deep learning for dialogue systems that have achieved good performance. They can generate fluent sentences based on a user’s utterances (Vinyals and Le, 2015; Shang et al., 2015; Serban et al., 2016). Nevertheless, such neural dialogue systems tend to generate phrases such as “Yes” and “I do not know” frequently in non-task-oriented dialog systems, referred to as the low diversity issue and the generic response issue. After training by a loss function of similarity with gold standard reference sentences, frequent phrases are more likely to be assigned a large occurrence probability than rare phrases are. Nakamura et al. (2018) proposed an Inverse Token Frequency (ITF) loss, which multiplies the Soft- max Cross-Entropy (SCE) loss by weights based on the inverse of the frequency of tokens. This ITF loss incorporates the frequency distribution of token classes so that rare tokens become more likely to appear. However, sentence diversity is based not only on individual tokens, but also on the token sequence. We are able to compute weights for loss functions dynamically, depending on the context, while retaining the fluency of generated sentences. We propose such a loss function, Inverse N-gram Frequency (INF) loss, which uses the inverse of the frequency of the n-gram of the tokens, rather than the token frequency. We built a neural dialogue system trained by INF loss using huge amounts of dialogue data extracted from Twitter. After comparing models using the SCE loss, the ITF loss, and the INF loss, we evaluated their diversity and fluency. Results show that our proposed INF loss model outperformed the SCE loss and ITF loss models for most automatic assessment measures such as DIST-N (Li et al., 2016) and ROUGE (Lin, 2004). Our INF loss model also achieved higher scores on our human evaluations of coherence and richness. 2 Related Work …

Content truncated to 500 words