Document clustering is an important tool, but itis not yet widely used in practice probably because of its high computational complexity. This article explores techniques of high-speed rough clustering of documents, assuming that it is sometimes necessary to obtain a clustering result in a shorter time, although the result is just an approximate outline of document clusters. A promising approach for such clustering is to reduce the number of documents to be checked for generating cluster vectors in the leader-follower clustering algorithm. Based on this idea, the present article proposes a modified Crouch algorithm and incomplete single-pass leaderfollower algorithm. Also, a two-stage grouping technique, in which the first stage attempts to decrease the number of documents to be processed in the second stage by applying a quick merging technique, is developed. An experiment using a part of the Reuters corpus RCV1 showed empirically that both the modified Crouch and the incomplete single-pass leader-follower algorithms achieve clustering results more efficiently than the original methods, and also improved the effectiveness of clustering results. On the other hand, the two-stage grouping technique did not reduce the processing time in this experiment.
Pseudo relevance feedback is empirically known as a useful method for enhancing retrieval performance. For example, we can apply the Rocchio method, which is well-known relevance feedback method, to the results of an initial search by assuming that the top-ranked documents are relevant. In this paper, for searching the NTCIR-3 patent test collection through pseudo feedback, we employ two relevance feedback mechanism; (1) the Rocchio method, and (2) a new method that is based on Taylor formula of linear search functions. The test collection consists of near 700,000 records including full text of Japanese patent materials. Unfortunately, effectiveness of our pseudo feedback methods was not empirically observed at all in the experiment.
Disambiguation between multiple translation choices is very important in dictionary-based cross-language information retrieval. In prior work, disambiguation techniques have used term co-occurrence statistics from the collection being searched. Experimentally these techniques have worked well but are based upon heuristic assumptions. In this paper, a theoretically grounded alternative is proposed, one which uses sense disambiguation based upon context terms within the source text. Specifically this paper introduces the concept of translation probabilities incorporating a context term and extends the IBM Model 1 for estimating context-based translation probabilities from a sentence-aligned bilingual corpus. Experimental results in English to Italian bilingual searches show significant performance improvement of the context-based translation probabilities over the case without any disambiguation.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.