Query Similarity of Various Linguistic Levels for Hybridized Conversational Agents

The performance of retrieval-based conversational agents is affected by the discrepancy between a user query and a retrieved query similar to the user query. There have been a number of previous studies to cope with this discrepancy, and a skeleton-based response generation is one of the successful approaches. However, it shows some ineffectiveness in that it considers only the lexical similarity in finding a similar query from a database of query-response pairs. Therefore, this paper proposes a CNN-based model which uses the combination of the neural representation of two queries and manually-designed lexico-syntactic features to determine the similarity between the queries. According to the experimental results on a manually-constructed dataset, the proposed model outperforms legacy search engine in finding similar queries from the database, which proves the plausibility of the proposed model.