Fluent Response Generation for Conversational Question Answering

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 191–207 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 191 Fluent Response Generation for Conversational Question Answering Ashutosh Baheti, Alan Ritter Computer Science and Engineering Ohio State University {baheti.3, ritter.1492}osu.edu Kevin Small Amazon Alexa smakevin@amazon.com Abstract Question answering (QA) is an important as- pect of open-domain conversational agents, garnering specific research focus in the con- versational QA (ConvQA) subtask. One no- table limitation of recent ConvQA efforts is the response being answer span extraction from the target corpus, thus ignoring the natural lan- guage generation (NLG) aspect of high-quality conversational agents. In this work, we pro- pose a method for situating QA responses within a SEQ2SEQ NLG approach to gener- ate fluent grammatical answer responses while maintaining correctness. From a technical per- spective, we use data augmentation to gen- erate training data for an end-to-end system. Specifically, we develop Syntactic Transforma- tions (STs) to produce question-specific can- didate answer responses and rank them using a BERT-based classifier (Devlin et al., 2019). Human evaluation on SQuAD 2.0 data (Ra- jpurkar et al., 2018) demonstrate that the pro- posed model outperforms baseline CoQA and QuAC models in generating conversational re- sponses. We further show our model’s scalabil- ity by conducting tests on the CoQA dataset.1 1 Introduction Factoid question answering (QA) has recently en- joyed rapid progress due to the increased availabil- ity of large crowdsourced datasets (e.g., SQuAD (Rajpurkar et al., 2016), MS MARCO (Bajaj et al., 2016), Natural Questions (Kwiatkowski et al., 2019)) for training neural models and the signifi- cant advances in pre-training contextualized repre- sentations using massive text corpora (e.g., ELMo (Peters et al., 2018), BERT (Devlin et al., 2019)). Building on these successes, recent work exam- ines conversational QA (ConvQA) systems capa- ble of interacting with users over multiple turns. 1The code and data are available at https://github.com/abaheti95/QADialogSystem. Large crowdsourced ConvQA datasets (e.g., CoQA (Reddy et al., 2019), QuAC (Choi et al., 2018)) consist of dialogues between crowd workers who are prompted to ask and answer a sequence of ques- tions regarding a source document. Although these ConvQA datasets support multi-turn QA interac- tions, the responses have mostly been limited to extracting text spans from the source document and do not readily support abstractive answers (Yatskar, 2019a). While responses copied directly from a Wikipedia article can provide a correct answer to a user question, they do not sound natural in a conversational setting. To address this challenge, we develop SEQ2SEQ models that generate fluent and informative answer responses to conversational questions. To obtain data needed to train these models, rather than constructing yet-another crowdsourced QA dataset, we transform the answers from an ex- isting QA dataset into fluent responses via data aug- mentation. Specifically, we synthetically generate supervised training data by converting questions and associated extractive answers from a SQuAD- like QA dataset into fluent responses via Syntactic Transformations (STs). These STs over-generate a large set of candidate responses from which a …

Content truncated to 500 words