Enhancing Dialogue Generation Models Via Multi Source Heterogeneous Information

Abstract

One of the fundamental goals in the field of Artificial Intelligence is to create machines newlinethat can interact with people using natural language, making research on virtual assistants newlineand dialogue systems a top priority for both industry and academia. Dialogue systems newlinehave emerged as a vital tool that helps people with their everyday tasks. Language newlineModeling and Natural Language Generation techniques have advanced rapidly, allowing newlinefor completely data-driven conversation models that produce natural language responses newlinedirectly. Building end-to-end dialogue models has been a long interest of natural language newlineresearch. However, it is a common problem that these dialogue generation models may newlinedevolve into boring and repetitive material, resulting in off-topic and pointless responses newlinein conversation assistants. In order to understand and generate coherent dialog, it is newlinenecessary to base it on extra information such as aspects, emotion, sentiment, images, and newlinestructured and unstructured knowledge. This thesis mainly focuses on enhancing Dialogue newlineGeneration Models via Multi-Source Heterogeneous Information in multiple domains. newlineEmotion plays an essential role in human communication, and it can significantly affect newlinethe content and tone of a conversation. However, incorporating emotion into dialogue newlinegeneration is challenging due to the subjective nature of emotions and the difficulty in newlinerepresenting them computationally. As part of this thesis, the first problem we investigate newlineis eliciting emotion in responses. We propose to build a deep multi-task framework that newlinecombines emotion classification and response generation for the task of emotion-controlled newlinedialog generation. Review-based knowledge refers to information extracted from user newlinereviews, which can be useful for providing personalized recommendations and answering newlinespecific questions. However, review data often contain noise and irrelevant information newlinethat can negatively impact the performance of dialogue systems. The second problem newlinewe investigate as part of this thesis