Term selection in information retrieval

Systems trained on linguistically annotated data achieve strong performance for many language processing tasks. This encourages the idea that annotations can improve any language processing task if applied in the right way. However, despite widespread acceptance and availability of highly accurate parsing software, it is not clear that ad hoc information retrieval (IR) techniques using annotated documents and requests consistently improve search performance compared to techniques that use no linguistic knowledge. In many cases, retrieval gains made using language processing components, such as part-of-speech tagging and head-dependent relations, are offset by significant negative effects. This results in a minimal positive, or even negative, overall impact for linguistically motivated approaches compared to approaches that do not use any syntactic or domain knowledge. In some cases, it may be that syntax does not reveal anything of practical importance about document relevance. Yet without a convincing explanation for why linguistic annotations fail in IR, the intuitive appeal of search systems that ‘understand’ text can result in the repeated application, and mis-application, of language processing to enhance search performance. This dissertation investigates whether linguistics can improve the selection of query terms by better modelling the alignment process between natural language requests and search queries. It is the most comprehensive work on the utility of linguistic methods in IR to date. Term selection in this work focuses on identification of informative query terms of 1-3 words that both represent the semantics of a request and discriminate between relevant and non-relevant documents. Approaches to word association are discussed with respect to linguistic principles, and evaluated with respect to semantic characterization and discriminative ability. Analysis is organised around three theories of language that emphasize different structures for the identification of terms: phrase structure theory, dependency theory and lexicalism. The structures identified by these theories play distinctive roles in the organisation of language. Evidence is presented regarding the value of different methods of word association based on these structures, and the effect of method and term combinations. Two highly effective, novel methods for the selection of terms from verbose queries are also proposed and evaluated. The first method focuses on the semantic phenomenon of ellipsis with a discriminative filter that leverages diverse text features. The second method exploits a term ranking algorithm, PhRank, that uses no linguistic information and relies on a network model of query context. The latter focuses queries so that 1-5 terms in an unweighted model achieve better retrieval effectiveness than weighted IR models that use up to 30 terms. In addition, unlike models that use a weighted distribution of terms or subqueries, the concise terms identified by PhRank are interpretable by users. Evaluation with newswire and web collections demonstrates that PhRank-based query reformulation significantly improves performance of verbose queries up to 14% compared to highly competitive IR models, and is at least as good for short, keyword queries with the same models. Results illustrate that linguistic processing may help with the selection of word associations but does not necessarily translate into improved IR performance. Statistical methods are necessary to overcome the limits of syntactic parsing and word adjacency measures for ad hoc IR. As a result, probabilistic frameworks that discover, and make use of, many forms of linguistic evidence may deliver small improvements in IR effectiveness, but methods that use simple features can be substantially more efficient and equally, or more, effective. Various explanations for this finding are suggested, including the probabilistic nature of grammatical categories, a lack of homomorphism between syntax and semantics, the impact of lexical relations, variability in collection data, and systemic effects in language systems