Word co-occurrence information utilization method for evaluated number independent retrieval on BBS articles with user evaluation

Recently, BBS (Bulletin Board System) has been used widely as online space to exchange information. It has a lot of useful information including users´ real voice such as review of new product, event information, and so on. However, sufficiently efficient functions for information search are not yet available to everyday users. BBS articles posted by general public have a lot of colloquial expressions and various expression patterns of same word. Therefore common searching methods, keyword matching and cosine similarity searching, can extract a lot of articles users do not want. As the result, it requires a fair amount of time for users to see if retrieved articles are necessary for them. The achievement of more efficiently retrieval function has been desired. A retrieval method using users´ evaluations of articles has been proposed. In this method, search user evaluates each retrieved thread on binary of whether its content suits the retrieval question. And when a new retrieval query is entered, the past evaluations by retrieval queries similar to the new query are used as feedback. It can improve retrieval accuracy on BBS browsed by many users and updated frequently, because a lot of evaluations can be stored. When there are few evaluations, however, this method makes a false selection of evaluations used as feedback. To solve this problem, in this paper, we propose a feedback method using users´ evaluations for evaluated number independent retrieval. In our method, appropriate feedback is achieved by using word co-occurrence information in articles and making word statistic classes, even if there are few evaluations. By using our method, the number of correct articles in retrieved result was increased compared to the conventional one when there are few evaluations.