Knowledge-aware response selection with semantics underlying multi-turn open-domain conversations

Response selection is a critical issue in the AI community, with important applications on the Web. The accuracy of the selected responses, however, tends to be insufficient due to the lack of contextual awareness, especially in open-domain conversations where words tend to have several meanings in different contexts. Our solution, SemSol, is a knowledge-aware response selection model that tackles this problem by utilizing the context-specific semantics behind words that are implicitly shared among users throughout the dialogue. SemSol simultaneously learns word sense disambiguations (WSD) for the words in the dialogue on the basis of an open-domain knowledge graph, i.e. WordNet, while learning the match between the context and the response candidates. Then, SemSol improves the accuracy of the response by exploiting the semantic information in a knowledge graph in accordance with the dialogue context. Our model learns the topics of utterances in the context of the whole training dataset. This topic-level knowledge can provide topic-specific information in the dialogue context. This improves the WSDs and the response selection accuracy. Experiments with two open-domain conversational datasets, Douban (Chinese) and Reddit (English), demonstrated that the SemSol model outperformed state-of-the-art baselines. SemSol is ranked #1 on the Douban leaderboard.