Discourse-Based Approach to Involvement of Background Knowledge for Question Answering

Proceedings of Recent Advances in Natural Language Processing, pages 373–381, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_044 373 Abstract We introduce a concept of a virtual discourse tree to improve question answering (Q/A) recall for complex, multi- sentence questions. Augmenting the discourse tree of an answer with tree fragments obtained from text corpora playing the role of ontology, we obtain on the fly a canonical discourse representation of this answer that is independent of the thought structure of a given author. This mechanism is critical for finding an answer that is not only relevant in terms of questions entities but also in terms of inter- relations between these entities in an answer and its style. We evaluate the Q/A system enabled with virtual discourse trees and observe a substantial increase of performance answering complex questions such as Yahoo! Answers and www.2carpros.com. 1 Introduction In spite of the great success of search technologies, the problem of leveraging background knowledge is still on the agenda of search engineering, for both conventional and learning-based systems. Background knowledge ontologies are difficult and expensive to build, and knowledge graphs – based approaches usually have a limited expressiveness and coverage. In this study we explore how a discourse analysis (which is domain- independent) can substitute certain features of ontology-based search. There are few popular discourse theories describing how Discourse Trees (DT) can be constructed from the text. In our work we used Rhetorical Structure Theory (RST, Mann and Thompson, 1988). Ontologies are in great demand for answering complex, multi-sentence questions with a precise answer in such domain as finance, legal, and health. In the educational domain this type of questions is referred to as convergent: answers to these types of questions are usually within a very limited range of acceptable accuracy. These may be at several different levels of cognition including comprehension, application, analysis, or ones where the answerer makes inferences or conjectures based on material read, presented or known. Answering convergent questions is an underexplored Q/A domain that can leverage discourse analysis (Kuyten et al, 2015). Discourse trees have became a standard for representing how thoughts are organized in text, in particular in a paragraph of text, such as an answer. Discourse-level analysis has been shown to assist in a number of NLP tasks where learning linguistic structures is essential (Louis et al., 2010; Lioma et al., 2012). DTs outline the relationship in between entities being introduced by an author. Obviously, there are multiple ways the same entities and their attributes are introduced, and not all rhetoric relations that hold between these entities occur in a DT for a given paragraph. When DTs are used to coordinate questions and answers, we would want to obtain an “ideal” DT for an answer, where all rhetoric relations between involved entities occur. To do that, we need to augment an actual (available) DT of answer instance with a certain rhetorical relations which are missing in the given answer instance but can be mined from text corpora or from the …

Content truncated to 500 words