Exploring and measuring dependency trees for informationretrieval

• Authors: Chang Liu
• Venue: Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval
• Year: 2008

Natural language processing techniques are believed to hold a tremendous potential to supplement the purely quantitative methods of text information retrieval. This has led to the emergence of a large number of NLP-based IR research projects over the last few years, even though the empirical evidence to support this has often been inadequate. Most contributions of NLP to IR mainly concentrate on document representation and compound term matching strategies. Researchers have noted that the simple term-based representation of document content such as vector representation is usually inadequate for accurate discrimination. The "bag of words" representation does not invoke linguistic considerations and allow modelling of relationships between subsets of words. However, even though a variety of content indicator such as syntactic phrase have been tried and investigated for representing documents rather than single terms in IR systems, the matching strategy over those representation still cannot go beyond traditional statistical techniques that measure term co-occurrence characteristics and proximity in analyzing text structure. In this paper, we propose a novel IR strategy (SIR) with NLP techniques involved at the syntactic level. Within SIR, documents and query representation are built on the basis of a syntactic data structure of the natural language text - the dependency tree, in which syntactic relationships between words are identified and structured in the form of a tree. In order to capture the syntactic relations between words in their hierarchical structural representation, the matching strategy in SIR upgrades from the traditional statistical techniques by introducing a similarity measure method executing on the graph representation level as the key determiner. A basic IR experiment is designed and implemented on the TREC data to evaluate if this novel IR model is feasible. Experimental results indicate that this approach has the potential to outperform the standard bag of words IR model, especially in response to syntactical structured queries.