Linguistically-Based Deep Unstructured Question Answering

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 433–443 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 433 Linguistically-based Deep Unstructured Question Answering Ahmad Aghaebrahimian Charles University Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostranske nam. 25, 11800 Praha 1, Czech Republic ebrahimian@ufal.mff.cuni.cz Abstract In this paper, we propose a new linguistically- based approach to answering non-factoid open-domain questions from unstructured data. First, we elaborate on an architecture for textual encoding based on which we in- troduce a deep end-to-end neural model. This architecture benefits from a bilateral attention mechanism which helps the model to focus on a question and the answer sentence at the same time for phrasal answer extraction. Second, we feed the output of a constituency parser into the model directly and integrate linguistic constituents into the network to help it concen- trate on chunks of an answer rather than on its single words for generating more natural out- put. By optimizing this architecture, we man- aged to obtain near-to-human-performance re- sults and competitive to a state-of-the-art sys- tem on SQuAD and MS-MARCO datasets re- spectively. 1 Introduction Reading, comprehending and reasoning over texts and answering a question about them (i.e. Ques- tion Answering) is a fundamental aspect of com- putational intelligence. Question Answering (QA), as a measure of intelligence, has been even suggested to replace Turing test (Clark and Et- zioni, 2016). The development of large datasets of QA in recent years (Hermann et al., 2015; Hill et al., 2015; Bordes et al., 2015; Rajpurkar et al., 2016) advanced the field especially for two significant branches of QA namely factoid (Aghaebrahimian and Jurˇc´ıˇcek, 2016a,b) and non-factoid QA1 (Ra- jpurkar et al., 2016). Non-factoid QA or QA over unstructured data is a somewhat new challenge in 1While the answer to a non-factoid question is a chunk of one or more adjacent words, the answer to a factoid question is only an entity. open-domain QA. A non-factoid QA system an- swers questions by reading and comprehending a context. The context in which we assume the an- swer is mentioned may have different granulari- ties from a single sentence or paragraph to larger units of text. A QA system is supposed to extract a phrase answer from the provided paragraph or sentence depending on its granularity level. The context for answering questions is usually extracted using an Information Retrieval (IR) tech- nique. Then, a QA system should extract the best answer sentence. There are many studies about ex- tracting answer sentences including but not limited to (He et al., 2015; He and Lin, 2016; Yih et al., 2013; Yu et al., 2016; Rao et al., 2016; Aghae- brahimian, 2017a). Extracting the final or shortest possible answer from a set of candidate answer sentences is ad- dressed in many studies as well (Zhang et al., 2017; Gong and Bowman, 2017; Shen et al., 2016; Weissenborn et al., 2017a). Instead of reasoning over and making inference on …

Content truncated to 500 words