Parse Thicket Representation for Multi-sentence Search

We develop a graph representation and learning technique for parse structures for sentences and paragraphs of text. This technique is used to improve relevance answering complex questions where an answer is included in multiple sentences. We introduce Parse Thicket as a sum of syntactic parse trees augmented by a number of arcs for inter-sentence word-word relations such as coreference and taxonomic. These arcs are also derived from other sources, including Rhetoric Structure theory, and respective indexing rules are introduced, which identify inter-sentence relations and joins phrases connected by these relations in the search index. Generalization of syntactic parse trees (as a similarity measure between sentences) is defined as a set of maximum common sub-trees for two parse trees. Generalization of a pair of parse thickets to measure relevance of a question and an answer, distributed in multiple sentences, is defined as a set of maximal common sub-parse thickets. The proposed approach is evaluated in the product search domain of eBay.com, where user query includes product names, features and expressions for user needs, and the query keywords occur in different sentences of text. We demonstrate that search relevance is improved by single sentence-level generalization, and further increased by parse thicket generalization. The proposed approach is evaluated in the product search domain of eBay.com, where user query includes product names, features and expressions for user needs, and the query keywords occur in different sentences of text.