Conversational Question Answering over Knowledge Graphs with Answer Verbalization

Abstract

In recent years, publicly available knowledge graphs (KG) have been broadly adopted as a source of knowledge in several tasks such as entity linking, relation extraction, and question answering. Question answering (QA) over knowledge graphs (KG) is an essential task that maps a user's utterance to a query over a KG to retrieve the correct answer. The initial knowledge graph question answering (KGQA) systems were mostly template or rule-based systems with limited learnable modules. Existing research on KGQA has accomplished outstanding results over simple questions, and lately, we have seen a successful effort to improve KGQA for complex questions. However, information needs are not always satisfied in one-shot processing. Either cause the first request is not well-formed, or the user requires more information for a particular topic and issues a series of follow-up questions. In such a conversational scenario, the follow-up questions are often incomplete since they co-reference entities and/or relations from previous interactions. Furthermore, existing KGQA resources and systems do not support answer verbalization. They provide nondescriptive answers extracted from KGs without verbalizing them in natural language utterances. <br> In this thesis, Conversational Question Answering over Knowledge Graphs with Answer Verbalization, we address conversational question answering and answer verbalization tasks while employing knowledge from structured graph data such as knowledge graphs. We present novel approaches based on deep neural network architectures and the multi-task learning paradigm, which allows for improved generalization. First, we propose extending question-answering resources with multiple verbalized answers to study whether the answer verbalization performance can be improved. In this work, we release the first question answering dataset with up to eight paraphrased responses for each question. We provide evaluation baselines to determine our dataset's quality and analyze the performance of various models when trained with one or more paraphrased answers. Next, we explore whether an additional source of information can be incorporated to generate better verbalized answers. Here, we develop a multi-task learning framework that comprises logical forms as auxiliary context alongside the questions. We evaluate on three QA datasets with answer verbalization where results establish a new baseline for the task. Afterward, we continue with the conversational question answering task, where we propose two multi-task neural semantic parsing approaches that construct logical forms and execute them in a knowledge graph to retrieve answers. We present an architecture of stacked pointer networks and we propose employing a Transformer model with Graph Attention Networks for consolidating knowledge graph information. We empirically study the proposed architectural design choices through an extensive evaluation and multiple analyses. Finally, we investigate the impact of incorporating answer verbalization in the conversational question answering task. We present an approach that jointly models the conversational context (entire dialog history with verbalized answers) and knowledge graph paths in a common space for learning joint embedding representations to improve path ranking. Results on standard datasets show a considerable improvement over previous baselines. Our contributions target a broader research agenda by providing efficient, conversational question answering and answer verbalization approaches. All the proposed approaches and resources presented in this thesis are publicly available.