Okapi Chinese Text Retrieval Experiments at TREC-6

The focus of the Okapi TREC{6 Chinese experiments is on investigating the e ectiveness of di erent automatic indexing methods and phrase weighting for retrieval based on probabilistic models over Chinese text. We compare di erent probabilistic weighting methods based on a range of word and single character approaches. There are two indexing methods used in our experiments. One indexing method is to use linguistic units (words, compound words and phrases) in texts as indexing terms to represent the texts. We refer to this method as the word approach. For this approach, text segmentation, which divides text into linguistic units, is regarded not only as a necessary precursor but also as a bottleneck of this kind of system [1]. The other method for indexing texts is based on single Chinese characters, in which texts are indexed by the characters appearing in the texts [2]. By using single character approaches, a search could be conducted for any multi-character word or phrase identi ed at search time, no matter whether this word or phrase is in the dictionary. Three automatic runs city97c1, city97c2 and city97c3 were submitted in TREC{6. All the three runs were based on the whole topic. City97c1 and city97c3 are for word indexing approach with di erent parameter values and city97c2 is for character indexing approach. The runs reported here are all on the TREC{6 collection of 26 new Chinese topics and 164768 documents. The Chinese dictionary we use for our word approach retrieval system contains about 70,000 Chinese words and phrases. Most of these words and phrases come from a manually constructed dictionary in China. We expanded this dictionary while working on the Chinese TREC experiments.