Finding Answers in Large Collections of Texts: Paragraph Indexing + Abductive Inference

This paper describes a methodology of answering questions by using information retrieved from very large collections of texts. We argue that combinations of information retrieval and extractions techniques cannot be used, due to the open-domain nature of the task. We propose a solution based on indexing techniques that identify paragraphs from texts where the answers can be found. The validity of the answers is obtained through a lightweight process of abduction. Background The process of answering a question starts with nding information from which the answer can be entailed. When the information is represented in axiomatic knowledge bases, logical resolution, refutation and theorem proving are the disciplines of AI that can help with the derivation of answers to questions posed in logical representations. In contrast, when the information is not represented in a formal language, but in natural language, answering questions has to deal with natural language ambiguities as well. Inf..