Towards Intelligent Text-Based Agents with Deep Learning

Nowadays, intelligent agents have gained a lot of attention and popularity due to the fast development of computing techniques and resources. Text-based intelligent agents can assist people in processing a vast amount of text information in their work and life. Several requirements for such agents need to be satisfied to effectively help people improve productivity, including understanding a single piece of text correctly, multi-document understanding, extracting/generating key concepts of a piece of text, and summarizing a long document. In this presentation, four research topics are discussed to help build those abilities for the intelligent text-based agents with deep learning methods. Firstly, I propose a neural network-based method utilizing information maximization to enhance the single-document understanding of the agent. Secondly, I design a hierarchical bi-directional self-attention network for multi-document understanding. Thirdly, I propose a joint learning method to generate key attribute values for a product from its textual title and description. Fourthly, I design a two-stage pipeline method to generate multiple aspect-based summaries for long meeting transcripts. In these research works, I only use public datasets including RCV1-V2, WOS, OpenReview, PeerRead, MEPAVE, AMI corpus and ICSI corpus.