Data Categorization Using Decision Trellises
IEEE Transactions on Knowledge and Data Engineering
| Hi-index | 0.01 |
Recently, BBS (Bulletin Board System) has frequently used because it contains a lot of users' valuable opinions. However, most users have some frustration when searching target information because BBS has lots of articles and there is no way to retrieve them efficiently and correctly. There are typical retrieval methods, so called "keyword matching" and "similarity-based search in word vector space", but neither of them provide desired retrieval results against incomplete or redundant sentences which are usually expressed in BBS articles. In order to solve this problem, we propose an efficient retrieval method that uses past other users' retrieval results. In our proposed method, users make two marks whether the retrieval result fits their desired one or not, and such marks are used as feedback for the similar retrieval of other users, which improves the correctness of retrieval results initially derived from "similarity-based search in word vector space". This framework does work well along with users' evaluation. We made an experiment by our proposed method, and the number of correct answers of ranking TOP10 of retrieval results increases about 50%, and the number of threads that users need to read by finding three desired articles is reduced from 10.6 to 3.2 on average compared with simply use of "similarity-based search in word vector space".