Subtopic Ranking Based on Block-Level Document Analysis

Abstract

We propose methods for ranking subtopics of a keyword query. Subtopics are also keyword queries which specialize and/or disambiguate search intent behind their original query. Information on subtopics are useful for search systems to generate diversified search results. Search result diversification is important when there are multiple ways to interpret the submitted query. In search result diversification, it is important to rank subtopics by their intent probabilities that users need information on the subtopics. Our subtopic ranking methods use hierarchical structure in documents in the corpus. Hierarchical structure in documents consists of nested logical blocks with headings. A heading describes the topic of a part of a document, and a block is such a part of a document. All our methods are based on two assumptions related to the structure. First, hierarchical headings in a document represent hierarchical topics discussed in the document. Second, authors write more contents about subtopics with higher intent probabilities. Based on these assumptions, our methods score each subtopic based on the total size of the blocks whose hierarchical headings represent the subtopic. We develop our methods in the following way. We first propose four methods to score a subtopic on a document, four methods to integrate subtopic scores on multiple documents, and two methods to sort subtopics based on their scores. We then combined these methods, which results in 32 subtopic ranking methods in total. We evaluated these methods on the data set for the subtopic mining subtask of the NTCIR-10 INTENT-2 task. The results indicated that our methods generated rankings statistically significantly better than the query completion snapshots by major commercial search engines.