Learn to Resolve Conversational Dependency: A Consistency Training Framework for Conversational Question Answering

Abstract

One of the main challenges in conversational question answering (CQA) is to resolve the conversational dependency, such as anaphora and ellipsis. However, existing approaches do not explicitly train QA models on how to resolve the dependency, and thus these models are limited in understanding human dialogues. In this paper, we propose a novel framework, EXCORD (Explicit guidance on how to resolve Conversational Dependency) to enhance the abilities of QA models in comprehending conversational context. EXCORD first generates self-contained questions that can be understood without the conversation history, then trains a QA model with the pairs of original and self-contained questions using a consistency-based regularizer. In our experiments, we demonstrate that EXCORD significantly improves the QA models' performance by up to 1.2 F1 on QuAC , and 5.2 F1 on CANARD (Elgohary et al., 2019), while addressing the limitations of the existing approaches. 1 † Corresponding author 1 Our models and code are available at: https://github.com/dmis-lab/excord 2 While the term "context" usually refers to the evidence document from which the answer is extracted, in CQA, it refers to conversational context.

Conversational Information Seeking

“…They focus on improving the contextualized encoding (BERT vs RoBERTa), multi-task learning of discourse or token importance, stacking networks to capture cross-turn relationships, and approaches to make the models more robust using adversarial training and data augmentation. Recent work by Kim et al (2021) brought together generative conversational query rewriting using T5 in the QA process and showed that it outperforms more complex models that attempt to model both simultaneously. The models largely target factoid QA with most being extractive, possibly with minor adaptions for yes/no questions or multiple choice.…”

Saving Dense Retriever from Shortcut Dependency in Conversational Search

“…For tackling the task, most works focused on developing model structures (Zhu et al, 2018;Qu et al, 2019;Zhao et al, 2021) or training strategies to encode the conversation history effectively. After the advent of a dataset containing standalone questions generated by human annotators (Elgohary et al, 2019), CQR approaches have been studied as a promising method for CQA Kim et al, 2021).…”

Towards More Realistic Generation of Information-Seeking Conversations

“…A line of research in conversational question generation (CQG) aims to generate humanlike follow-up questions upon conversational history Pan et al, 2019;Qi et al, 2020;Gu et al, 2021). Another line of research has greatly improved answer accuracy Qu et al, 2019b;Kim et al, 2021;Zhao et al, 2021). In other words, they are limited in assuming that all other ingredients (i.e., held-out conversations by humans and their gold answer) are provided.…”

“…Elgohary et al (2019) rewrite conversational questions of QuAC into self-contained questions that could be understood without the conversation. Following Kim et al (2021), we consider the resulting dataset, CANARD, as an additional dataset for training CQA models. Note that CANARAD train and QuAC seen share the same passages.…”

“…Several works demonstrated the effectiveness of the flow mechanism in CQA . Most recently, leveraging self-contained questions (Kim et al, 2021) or encoding longer context (Zhao et al, 2021) have been shown to be effective in the task.…”