Diverse and Informative Dialogue Generation with Context-Specific Commonsense Knowledge Awareness

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5811–5820 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5811 Diverse and Informative Dialogue Generation with Context-Specific Commonsense Knowledge Awareness Sixing Wu1 , Ying Li2∗, Dawei Zhang1 , Yang Zhou3 and Zhonghai Wu2 1School of Electronics Engineering and Computer Science, Peking University, Beijing, 100871, China 2National Research Center of Software Engineering, Peking University, Beijing, 100871, China 3Auburn University, Auburn, Alabama, 36849, USA {wusixing, li.ying, daweizhang}@pku.edu.cn yangzhou@auburn.edu, zhwu@ss.pku.edu.cn Abstract Generative dialogue systems tend to produce generic responses, which often leads to bor- ing conversations. For alleviating this issue, Recent studies proposed to retrieve and intro- duce knowledge facts from knowledge graphs. While this paradigm works to a certain ex- tent, it usually retrieves knowledge facts only based on the entity word itself, without consid- ering the specific dialogue context. Thus, the introduction of the context-irrelevant knowl- edge facts can impact the quality of genera- tions. To this end, this paper proposes a novel commonsense knowledge-aware dialogue gen- eration model, ConKADI. We design a Felic- itous Fact mechanism to help the model fo- cus on the knowledge facts that are highly rel- evant to the context; furthermore, two tech- niques, Context-Knowledge Fusion and Flex- ible Mode Fusion are proposed to facilitate the integration of the knowledge in the ConKADI. We collect and build a large-scale Chinese dataset aligned with the commonsense knowl- edge for dialogue generation. Extensive eval- uations over both an open-released English dataset and our Chinese dataset demonstrate that our approach ConKADI outperforms the state-of-the-art approach CCM, in most exper- iments. 1 Introduction Nowadays, open-domain dialogue response gen- eration systems have shown impressive poten- tial, to endow a machine with the ability to con- verse with a human, using natural language (Chen et al., 2017). Although such models have achieved promising performance, they still suffer from gen- erating generic and boring responses, such as ”I don’t know.” Such low-quality responses always reduce the attractiveness of generative dialogue systems to end-users. Researchers have tried to ∗Corresponding author: Ying Li, li.ying@pku.edu.cn tackle it from multiple aspects; for example, using the enhanced objective function (Li et al., 2016a); introducing additional contents (Xu et al., 2019). However, these methods haven’t solved the issue thoroughly. Different from a human being, who is capable of associating the dialogue with the back- ground knowledge in his/her mind, a machine can merely capture limited information from the sur- face text of the query message (Ghazvininejad et al., 2018). Consequently, it is difficult for a machine to understand the query fully, and then to gener- ate diverse and informative responses (Zhou et al., 2018). To bridge the gap of the knowledge between the human and the machine, researchers have begun to introduce large-scale knowledge graphs for en- hancing the dialogue generation (Zhu et al., 2017; Zhou et al., 2018; Liu et al., 2018), and they have obtained lots of impressive results. Generally, the retrieval of knowledge facts is based on the entity name; in …

Content truncated to 500 words