A Ranking Method for Relaxed Queries in Book Search

In this paper, we propose a ranking method for keyword-based book search systems. Because we do not have full text data of all books, we use a database of brief descriptions of books. Such a brief description of the target book may not include all the query keywords given by the user. In addition, a query formulated by a user based on a vague memory may include wrong keywords. In both cases, the target book does not match with the query. To solve that problem, our method ranks candidate books in two steps. We first generate relaxed queries by removing some keywords from the given original query, and rank the queries based on how likely the remaining keywords are to appear in the brief descriptions. In the second step, we rank the results of each query. For each query, we retrieve matching books, find words in the description of each book that are most likely to be mistaken for the removed keywords by the user, and rank the books based on how likely they are mistaken for the removed keywords. By combining these two rankings, i.e., the ranking of relaxed queries, and the ranking of books matching with each query, we produce the final ranking. In this paper, we focus on the ranking method for the second step. Our experiment shows that our method improves ranking for some queries where the original query includes many keywords that do not appear in the description of the target book.