Enriching Text-Based Human-Machine Interactions with Additional World Knowledge

Abstract

In the last decade, the world has witnessed a massive upheaval in intelligent systems' design and development, especially for systems that can interact with humans. Such systems, capable of real-time interaction with end users can be realized in several ways, ranging from text/speech based to gesture or electroencephalogram (EEG). Speech/text-based human-machine interactive (HMI) systems have become part of human's everyday life in the form of Alexa, Siri, etc. Aforesaid HMI systems can be realized as question-answering systems or Dialogue systems/chatbots; where the former is capable of answering individual queries from the user while the latter has an additional understanding of the conversational context. <br /> Both kinds of systems can potentially benefit from additional world knowledge, which is present across both unstructured and structured sources. Unstructured sources consist of textual knowledge present in the form of definitions or text articles; while, structured knowledge can be present in the form of, e.g., Knowledge Graphs or RDBMS. However, integrating both structured and unstructured knowledge while designing such systems is challenging because the models have to solve multiple tasks of inferencing, coherent response generation, and knowledge integration. <br /> In this thesis, we explore different techniques for integrating additional world knowledge into methods used for implementing question-answering and dialogue systems. Firstly, we examine different techniques to incorporate structural information present in a Knowledge Graph (KG) into Knowledge-Graph-based question answering systems (KGQA). Here, we investigate how to improve KGQA with added structural information of the Knowledge Graph. Most KGQA systems do entity and relation linking separately and have no deeper understanding of the connected KG components and structure of the KG elements. We explore and propose different algorithms for KG-aware question answering, both as modular and end-to-end settings. Secondly, we investigate different techniques for incorporating world knowledge, present in the form of both structured and unstructured sources into multi-turn conversational systems. For unstructured knowledge integration, we probe end-to-end techniques to make retrieval-based chatbots better using textual definitions of in-domain keywords. Henceforth, we explore techniques to incorporate structured information present in the form of Knowledge Graphs into generative dialogue systems. Several novel neural network architectures are proposed which implicitly look into the Knowledge Graph during the dialogue generation process. Some architectures are for handling goal-oriented dialogues while others are designed to handle both goal and non-goal oriented dialogues. For the first model, we incorporate joint Knowledge Graph and word embeddings, projected in the same vector space for task-oriented dialogues. Secondly, we incorporate a copy mechanism into the response generation process for end-to-end generative dialogue systems, where the model can probabilistically copy information from the underlying Knowledge Graph-based on the question. Finally, we use a pre-trained transformer (BERT) model which is capable of answering from the Knowledge Graph by jointly performing entity linking, relation identification while generating coherent responses at the same time.