Abstract: In an active cyberspace ecosystem, Natural Language Processing (NLP) algorithms and tasks are essential components. The massive text content is often accompanied with abundant heterogeneous data, e.g., click logs, knowledge graphs, chronological search history, and computational user profiles, which provide us opportunities of taking advantage of data heterogeneity for further advancing NLP algorithms. By leveraging various kinds of user-generated data and sophisticated deep learning methodologies, we proposed a number of algorithms and applications to address different novel and classical research questions, and, moreover, we also successfully validated and adopted them on industry platforms to serve millions of users. In this talk, I will present three cases: 1) Cybercrime Detection with Berrypicking Behaviors, 2) Natural Language Generation (NLG) with Click-logs, and 3) Multi-role Dialogue Mining with Multi-task Pretraining. In these works, user-centric heterogeneous data and text information are jointly encapsulated by using innovative representation learning and multi-task learning.

Bio: Xiaozhong Liu is an Associate Professor from School of Informatics, Computing and Engineering at Indiana University Bloomington. He also serves as senior consultant to Alibaba DAMO Academy, the research arm of Alibaba, and leads large-scale NLP projects. He has published more than 150 papers in leading computer science conferences and information science journals, e.g, PNAS, JASIST, AAAI, SIGIR, IJCAI, EMNLP, ACL and WWW, and holds nine patents in AI and NLP. His areas of research interest include Data Science, NLP, Explainable AI, Graph Mining, Cybercrime and Security, and Computational Social Science. Currently, his algorithm APIs are called more than 170 million times per day on different active eCommerce and Web Search platforms and effectively outperforms predecessor methods by 10% on average.

RSVP for Zoom information: https://pitt.co1.qualtrics.com/jfe/form/SV_3OGembQH1nlTHcG