Natural Language Information Retrieval

Information Retrieval (IR) in this collection of 14 original papers is broadly construed to include document retrieval, information extraction, question answering, clustering and classification. In the introduction Strzalkowski asks the provocative question Why hasn't NLP had more success in IR?, the answer to which ought to be of interest to readers of Computational Linguistics. Unfortunately, the majority of the papers in the book completely fail to address this question, and several of the papers do not discuss applications of NLP to IR at all. A brief summary of the papers demonstrates the range of topics covered. Karen Sparck Jones, in "What is the role of NLP in text retrieval?," gives an overview of linguistically motivated indexing (LMI) and nonlinguistic indexing (NLI). LMI is not suitable for queries with few words, yet as more words are added, the conjunction of search terms benefits NLI, raising the bar against which LMI is being compared. Sparck Jones concludes that LMI might still be useful for displaying informative information about documents. Christian Jacquemin and Evelyne Tzoukermann, in "NLP for term variant extraction: Synergy between morphology, lexicon, and syntax," perform phrase normalization on the basis of full morphological analysis and patterns over parts of speech and syntactic constituents. They provide an overview of finite-state automata for morphological analysis and rule ordering for derivational affixation in French, with a tangential section on Spanish. Gerda Ruge, in "Combining corpus linguistics and human memory models for automatic term association," draws on psycholinguistic research to improve models of spreading activation within a semantic network sensitive to head/modifier relationships. Alan F. Smeaton, in "Using NLP or NLP resources for information retrieval tasks," after experiments with matching entire syntactic analyses for TREC yielded results that were much worse than traditional tf.idf measures, has experimented with selectively using NLP resources such as WordNet. Retrieval of picture captions with manual word sense disambiguation outperforms a tf.idf baseline. Tomek Strzalkowski, Fang Lin, Jin Wang, and Jose Perez-Carballo, in "Evaluating natural language processing techniques for information retrieval: A TREC perspective," search for appropriate ways to weight linguistic and nonlinguistic representations of document content, and explore expansions of the query based on selecting entire paragraphs. Manually selected paragraphs yield substantial gains in precision.