ExtrAns - Answer Extraction from Technical Documents by Minimal Logical Forms and Selective Highlighting

Abstract

Logic-based answer extraction techniques present a solution to retrieve and mark those exact passages in a document that directly answer a natural language query. In contrast to pure information retrieval techniques that treat content words as isolated terms, answer extraction techniques exploit syntactic information in a document to a certain degree and consider semantic relations between function words and content words. Minimal logical forms (MLF) - specially designed for this task - represent the semantic relations of the sentences and point to the textual information in the document. MLFs consist of existentially closed atomic formulas and use reification of objects, eventualities and properties as a building principle. On account of their simple design MLFs proved to be computationally tractable and incrementally extensible in our answer extraction system ExtrAns. Unresolved structural ambiguities are represented by alternative MLFs. The theorem prover of ExtrAns finds all proofs for an (ambiguous) query and considers the frequency of a part of a MLF used during the proof as an indicator for the retrieval relevance. The actual retrieval relevance is reflected by selective highlighting in the document. The more often a part of a MLF that points to a specific phrase of a sentence is used for the proof, the more intensively this phrase is marked by the colouring scheme.