Towards More Viable Natural Language Interaction

Abstract

Making machines interact viably with humans in natural language is part of the most elusive tasks to realize Artificial Intelligence, where machines need to understand the language from humans and make proper responses. This thesis targets two important types of problems related to language interaction, Question Answering (QA) and Dialogue Generation. The former requires the machine to take a question from humans and provide the correct answer by understanding the question as well as related knowledge such as supporting contexts. In the second task, the machine needs to generate proper textual responses conditioned on the dialogue history with humans, possibly along with auxiliary information, to finish a conversation. There are two main parts of the thesis. In the first part, we aim to solve QA tasks. We first propose BAG, a Bidirectional Attention Entity Graph Convolutional Network, for multi-hop QA tasks. Next, we move to the robustness of current models. We first study the transferability of QA models from one dataset to another one, based on the large performance divergence between different datasets. Then we attack QA models adversarially, finding the pitfalls and designing corresponding adversarial samples to alleviate these side effects. In the second part, we move to dialogue generation systems. Our first contribution is a model integrated with pre-trained models for multi-input dialogue generation tasks, such as personalized dialogue where responses need to reflect the given personas. Our second contribution is data augmentation approaches to enhance the performance. We propose an embedding-level mixup augmentation for general purpose, along with a data manipulation method, especially for personalized dialogue tasks, including data distillation, data diversification, and data curriculum. Our works show sufficient superiority to existing ones, and they can hold great promise for future language technologies.