PIQASso: PIsa Question Answering System

PiQASso is a Question Answering system based on a combination of modern IR techniques and a series of semantic filters for selecting paragraphs containing a justifiable answer. Semantic filtering is based on several NLP tools, including a dependency-based parser, a POS tagger, a NE tagger and a lexical database. Semantic analysis of questions is performed in order to extract keywords used in retrieval queries and to detect the expected answer type. Semantic analysis of retrieved paragraphs includes checking the presence of entities of the expected answer type and extracting logical relations between words. A paragraph is considered to justify an answer if similar relations are present in the question. When no answer passes the filters, the process is repeated applying further levels of query expansions in order to increase recall. We discuss results and limitations of the current implementation. 1. Architecture The overall architecture of PiQASso is shown in Figure 1 and consists in two major components: a paragraph indexing and retrieval subsystem and a question answering subsystem. The whole document collection is stored in the paragraph search engine, through which single paragraphs are retrieved, likely to contain an answer to a question. Processing a question involves the following steps: • question analysis • query formulation and paragraph search • answer type filter • relation matching filter • popularity ranking • query expansion. Question analysis involves parsing the question, identifying its expected answer type and extracting relevant keywords to perform paragraph retrieval. The initial query built with such keywords is targeted to high precision and to retrieve a small number of sentences to be evaluated as candidate answers through a series of