Web Document Retrieval Using Sentence-Query Similarity.
Eui-Kyu Park, Seong‐In Moon, Dong-Yul Ra, Myung-Gil Jang · 2002
For the web document retrieval experiments in our TREC '2002 participation, we used two new methods. One is the use of anchor texts, which has been advocated by many researchers. But the methods used by them is different from our method. The second is the use of sentence-query similarity. It has been known that the use of links for web retrieval did not show impressive improvement in performance [5,6,8,9]. But Bailey, etc. [1] reported that using anchor texts can improve retrieval performance. However, our home page finding experiment done for TREC '2001 showed that it is not the case. The use of anchor texts did not allow any improvement in performance. Our method to use the anchor texts this year is changed a lot from last year and found that it is pretty effective. The major focus of our experiment this year is in the use of sentential information in information retrieval. We obtain similarity values between sentences of a document and the query and use them for computing the retrieval score of the document. The main idea is the following: a sentence in a document that is much relevant to the query can support relevance of the document to the query. We compute the similarity between each sentence in the document and the query. The degree of this similarity is incorporated in calculating the document's score (in addition to the similarity between the document as a whole and the query). It has been found that it does not take too much time for this extra processing. Our experiment showed that including the sentential information in the proposed way can significantly improve retrieval effectiveness.