Improving Knowledge-Aware Dialogue Response Generation by Using Human-Written Prototype Dialogues

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1402–1411 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1402 Improving Knowledge-Aware Dialogue Response Generation by Using Human-Written Prototype Dialogues Sixing Wu1 , Ying Li2∗, Dawei Zhang1 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 {wusixing, li.ying, daweizhang}@pku.edu.cn zhwu@ss.pku.edu.cn Abstract Incorporating commonsense knowledge can alleviate the issue of generating generic re- sponses in open-domain generative dialogue systems. However, selecting knowledge facts for the dialogue context is still a challenge. The widely used approach Entity Name Match- ing always retrieves irrelevant facts from the view of local entity words. This paper pro- poses a novel knowledge selection approach, Prototype-KR, and a knowledge-aware gener- ative model, Prototype-KRG. Given a query, our approach first retrieves a set of prototype dialogues that are relevant to the query. We find knowledge facts used in prototype dia- logues usually are highly relevant to the cur- rent query; thus, Prototype-KR ranks such knowledge facts based on the semantic similar- ity and then selects the most appropriate facts. Subsequently, Prototype-KRG can generate an informative response using the selected knowl- edge facts. Experiments demonstrate that our approach has achieved notable improve- ments on the most metrics, compared to gen- erative baselines. Meanwhile, compared to IR(Retrieval)-based baselines, responses gen- erated by our approach are more relevant to the context and have comparable informativeness. 1 Introduction Unlike human beings, generative dialogue systems tend to generate generic responses, such as ‘I don’t know.’ (Li et al., 2016). One possible reason is the gap in utilizing background knowledge. Hu- man beings can naturally frame their dialogue un- derstanding and responding with various learned background knowledge during the conversation. However, traditional dialogue systems can merely access the surface knowledge in the given query (Ghazvininejad et al., 2018). To tackle this is- sue, a feasible scheme is incorporating external ∗Corresponding author. knowledge into the dialogue generation (Qin et al., 2019; Wu et al., 2020b). This paper focuses on introducing the structured open-domain common- sense knowledge graph into the single-turn dia- logue response generation. Commonsense knowl- edge refers to the widely-used everyday knowledge, for example, ‘lemon tastes sour’. In general, a knowledge graph can be regarded as a set of (ehead, r, etail) fact triplets. For the knowledge-aware dialogue generation, the first step is knowledge selection, aiming at selecting appro- priate knowledge facts for the current dialogue context. Traditional works (Zhou et al., 2018) al- ways adopt the Entity Name Matching (ENM), i.e., knowledge facts are retrieved based on the entity words that appear in the given query. For example, the fact triplet (apple, IsATypeOf, fruit) can be se- lected for the query ‘What’s your favourite fruit?’. Although such a widely-used method works to some extent, it has several flaws. First, only 1- hop knowledge can be retrieved. Second, instead of using the utterance-level (global) features, it uses local words to retrieve; thus, irrelevant knowledge facts …

Content truncated to 500 words