The Effect of Weighted Term Frequencies on Probabilistic Latent Semantic Term Relationships

Abstract

Probabilistic latent semantic analysis (PLSA) is a method of calculating term relationships within a document set using term frequencies. It is well known within the information retrieval community that raw term frequencies contain various biases that affect the precision of the retrieval system. Weighting schemes, such as BM25, have been developed in order to remove such biases and hence improve the overall quality of results from the retrieval system. We hypothesised that the biases found within raw term frequencies also affect the calculation of term relationships performed during PLSA. By using portions of the BM25 probabilistic weighting scheme, we have shown that applying weights to the raw term frequencies before performing PLSA leads to a significant increase in precision at 10 documents and average reciprocal rank. When using the BM25 weighted PLSA information in the form of a thesaurus, we achieved an average 8% increase in precision. Our thesaurus method was also compared to pseudo-relevance feedback and a co-occurrence thesaurus, both using BM25 weights. Precision results showed that the probabilistic latent semantic thesaurus using BM25 weights outperformed each method in terms of precision at 10 documents and average reciprocal rank.