Diversified and Verbalized Result Summarization for Semantic Association Search

Semantic association search is to search an entity-relation graph for subgraphs called semantic associations that connect a set of entities specified in a user’s query. Recent research on this topic has concentrated on summarizing numerous search results by mining their important patterns to form an abstractive overview. However, top-ranked patterns may have redundancy, and their graph structure may not be comprehensible to non-expert users. To reduce redundancy, we present a novel framework featuring a combinatorial optimization model to select top-k diversified patterns. In particular, we devise a new similarity measure which jointly considers structural and semantic similarity to assess the overlap between patterns. To facilitate non-expert users’ comprehension of a pattern, we verbalize its graph structure, transforming it into compact and coherent English text based on a novel method for discourse planning. Extensive experiments demonstrate the effectiveness of our approach compared with existing methods.