TREC-5 English and Chinese Retrieval Experiments using PIRCS.
Kui Lam Kwok, Laszlo Grunfeld, Jinyong Xu · 1996
For Trec-6 ad-hoc experiments, we continue to use twostage retrieval with pseudo-feedback from top-ranked unjudged documents for both Chinese and English. We perform three types of retrieval characterized by queries formed using title only, description only and all sections of the given topics. For short queries mainly derived from title or description section, query terms are weighted by average term frequency avtf introduced previously. For Chinese, we employ a combination of representation (character, bigram and short-word) strategy, returning the highest average non-interpolated precision that is even better than some manual approaches. In English ad-hoc, we try a document re-ranking strategy for the first stage retrieval based on occurrence of selected query term pairs, so as to have better result in the second stage. Performance for English ad-hoc is also highly competitive for both very short and long queries. In routing, a strategy of combining different methods of query format...