A Comparative Study of Probabilistic and Language Models for Information Retrieval

Language models for information retrieval have received much attention in recent years, with many claims being made about their performance. However, previous studies evaluating the language modelling approach for information retrieval used different query sets and heterogeneous collections, which make reported results difficult to compare. This research is a broad-based study that evaluates language models against a variety of search tasks — topic finding, named-page finding and topic distillation. The standard Text REtrieval Conference (TREC) methodology is used to compare language models to the probabilistic Okapi BM25 system. Using consistent parameter choices, we compare results of different language models on three different search tasks, multiple query sets and three different text collections. For ad hoc retrieval, the Dirichlet smoothing method was found to be significantly better than Okapi BM25, but for named-page finding Okapi BM25 was more effective than the language modelling methods. Optimal smoothing parameters for each method were found to be dependent on the collection and the query set. For longer queries, the language modelling approaches required more aggressive smoothing but they were found to be more effective than with shorter queries. The choice of smoothing method was also found to have a significant effect on the performance of language models for information retrieval.