Abstract: In this talk, I will present two applications on one of the most popular e-commerce platforms: Amazon. We will first look at generating product descriptions, and show how we train a multimodal model that takes product images and metadata as input to generate descriptions for 23 product types. Although we found that the generated descriptions are fluent and persuasive, they are not always faithful (i.e. the details are not real). Next I will introduce the task of product question answering, where the goal is to automatically generate answers given customer queries for a particular product. Here we explore mining answers from product reviews to answer queries, and propose a mixture-of-expert model that can be trained using only existing answered queries without needing annotated query-review data.
Bio: Dr Lau is a lecturer in the School of Computing and Information Systems at the University of Melbourne. His research is in Natural Language Processing — a sub-field of Artificial Intelligence — where the goal is to develop computational models to understand human languages. A common theme of Dr Lau's research is that it involves building computational models in an unsupervised or semi-supervised setting, i.e. a learning scenario where the supervision signal for model training is not available or scarce, and is characterised by a diverse flavour of applications, e.g. topic models, lexical semantics, text generation and misinformation detection. Some of his research in text generation and state-sponsored influence operations has been covered by popular science magazines (New Scientist) and mainstream news media (BBC and Guardian).