How Part-of-Speech Tags Affect Text Retrieval and Filtering Performance

Abstract

Natural language processing (NLP) applied to information retrieval (IR) and filtering problems may assign part-of-speech tags to terms and, more generally, modify queries and documents. Analytic models can predict the performance of a text filtering system as it incorporates changes suggested by NLP, allowing us to make precise statements about the average effect of NLP operations on IR. Here we provide a model of retrieval and tagging that allows us to both compute the performance change due to syntactic parsing and to allow us to understand what factors affect performance and how. In addition to a prediction of performance with tags, upper and lower bounds for retrieval performance are derived, giving the best and worst effects of including part-of-speech tags. Empirical grounds for selecting sets of tags are considered.