Keynote & Invited Talk
Title: Deep Learning for Natural Language Processing: Current and Future
Abstract:
Why is deep learning so powerful for natural language processing? How will deep learning for natural language processing evolve in the future? These are important questions which many researchers in the field may be thinking of. In this talk, I will provide my answers to the questions. First, I will summarize the advantages and disadvantages of deep learning, as well as the major problems of natural language processing. Next, I will discuss the reason behind the success of deep learning in natural language processing, particularly search and recommendation. Finally, I will share my view on future directions of deep learning for natural language processing.
Title: Explanatory Natural Language Processing: Formulation, Methods, and Evaluation.
Abstract: Building explanatory models and applications is widely considered as a critical component towards the fairness, accountability, and transparency of machine learning. In practice, however, the necessity and the practical value of explaining a black-box neural network model are still under debate. This controversy is largely due to the lack of a clean formulation and objective evaluation about the explainability of a model. In this talk, I will introduce a novel conceptual formulation of explanatory machine learning, which centers on how human users make joint decisions with a machine learning based predictor. This general framework also leads to a natural and robust way of evaluating the explanations, through which the goodness of an explanatory model is measured by how much better the joint decisions are with vs. without the explanations. I will introduce our recent work along this direction, including an end-to-end adversarial attention network for explanatory natural language processing, its application to identifying toxicity from social media posts, and user studies at scale to evaluate its effectiveness in practice.
Invited Talk
Title: Democratize conversational user interface
Abstract: Chatbot is essentially an application that allows users to interact with your business through a conversational interface. Due to its lower learning curve, and the ability to directly access information and services, chatbots are widely predicted to be the next-generation user interface and hence attracted attention from both research and industry. For example, two years in a row, Dialogue is the dominant topic at ACL. But despite the interest and effort level on chatbot, in reality, the good conversational user experience is still far and between. In this talk, we will analyze some common misconceptions about building conversational interface, hypothesize why building GUI app is so much cheaper/easier than building their CUI counterpart and discuss ways that are needed to make the conversational user interface as ubiquitous as everyday apps on your phone.