Question-Answering Using Semantic Relation Triples

This paper describes the development of a prototype system to answer questions by selecting sentences from the documents in which the answers occur. After parsing each sentence in these documents, databases are constructed by extracting relational triples from the parse output. The triples consist of discourse entities, semantic relations, and the governing words to which the entities are bound in the sentence. Database triples are also generated for the questions. Question-answering consists of matching the question database records with the records for the documents. The prototype system was developed specifically to respond to the TREC-8 Q&A track, with an existing parser and some existing capability for analyzing parse output. The system was designed to investigate the viability of using structural information about the sentences in a document to answer questions. The CL Research system achieved an overall score of 0.281 (i.e., on average, providing a sentence containing a correct answer as the fourth selection). The score demonstrates the viability of the approach. Post-hoc analysis suggests that this score understates the performance of the prototype and estimates that a more accurate score is approximately 0.482. This analysis also suggests several further improvements and the potential for investigating other avenues that make use of semantic networks and computational lexicology.