Two Discourse Tree - Based Approaches to Indexing Answers

Proceedings of Recent Advances in Natural Language Processing, pages 367–372, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_043 367 1 Two Discourse Tree - Based Approaches to Indexing Answers Boris Galitsky1 and Dmitry Ilvovsky2 1Oracle Inc. Redwood Shores CA 2National Research University Higher School of Economics boris.galitsky@oracle.com; dilvovsky@hse.ru Abstract We explore anatomy of answers with re- spect to which text fragments from an an- swer are worth matching with a question and which should not be matched. We ap- ply the Rhetorical Structure Theory to build a discourse tree of an answer and se- lect elementary discourse units that are suitable for indexing. Manual rules for se- lection of these discourse units as well as automated classification based on web search engine mining are evaluated con- cerning improving search accuracy. We form two sets of question-answer pairs for FAQ and community QA search domains and use them for evaluation of the pro- posed indexing methodology, which deliv- ers up to 16 percent improvement in search recall. 1 Introduction Much online content is available via question- answer pairs such as frequently-asked questions stored on customer portals or internal company portals. Question-answer pairs can be an efficient manner to familiarize a user with content. In some cases, autonomous agents (chatbots) can import such question-answer pairs in order to field user questions. But such question-answer pairs can contain content that is not central to a topic of an answer. For example, content can include text that is ir- relevant or misleading, non-responsive to the particular question, or is neutral and not helpful. If irrelevant text is indexed by a keyword-based search engine, the precision of the search engine is lowered. Moreover, an autonomous agent at- tempting to answer a user question based on er- roneously-indexed text may answer the question incorrectly, resulting in lowered user confidence in the agent. Despite the fact that standard rele- vance techniques such as ontology, keyword fre- quency models and separate discourse features (Chali et al., 2009; Jansen et al., 2014) can be applied to solve this problem, a solution is need- ed for identifying informative parts from all text. In this paper we propose a new discourse- based approach to determine informative parts of an answer. This approach accesses a body of text including fragments and creates a searchable in- dex including multiple entries, each entry corre- sponding to a selected fragment. We propose two different methods of fragment selection based on rules and on classification model respectively. The paper structure is as follows. In Section 2 we introduce the methodology of rhetorical anatomy of an answer and present an example of that. In Section 3 we propose two Q/A algo- rithms which is the core part of our approach. In Section 4 we describe and discuss evaluation for the question answering task on a few datasets that were compiled for this research. 2 Rhetoric Anatomy of an Answer 2.1 RST and Discourse Trees Discourse analysis was proved to be useful in different aspects of question-answering: answer extraction (Zong …

Content truncated to 500 words