2017-10-10 | Yu Su：On Building Natural Language Interface to Data and Services
On Building Natural Language Interface to Data and Services（PDF）
By mapping natural languages to the formal languages used by computers, natural language interface holds the promise for serving as a unified interface to seamlessly access and interact with a great variety of data and services. For decades people have long been envisaging generic natural language interfaces that grant all users, regardless of their proficiency in programming, equal access to the digital world. With the recent advances in deep learning and knowledge base, it is widely believed that we are better equipped than ever to tackle some most challenging problems in this area, and practical, well-functioning natural language interfaces for a wide range of domains are about to be born.
In this talk, I will present our recent explorations in natural language interface. The talk will start with discussion on a holistic design of practical natural language interfaces, covering important topics like semantic parsing, scalability, interpretability and interactivity. I will then specifically address one challenging question: Given a new domain with **zero** training data and user, how to build a natural language interface from scratch? We discuss two complementary solutions, one using crowdsourcing to collect training data at a low cost, and the other using existing data from other domains via neural transfer learning. Finally, I will briefly describe our studies towards making natural language interfaces more robust and usable, such as answer triggering, i.e., checking whether a question is answerable, and post-inspection, i.e., checking whether the parsing result is sensible before execution. I will end the talk with a few exciting future directions including cross-domain natural language interface and application to important vertical domains like healthcare and finance.
Yu Su is a final-year Ph.D. candidate in the Department of Computer Science at University of California, Santa Barbara. He works with Xifeng Yan on a variety of topics at the intersection of data mining, natural language processing, and machine learning. Before coming to UCSB, he got his bachelor degree from the Department of Computer Science and Technology at Tsinghua University in 2012. His recent research interests include natural language interface, knowledge base construction, and general text mining, with a focus on understanding the interplay of natural and formal languages to increase the accessibility of data (e.g., knowledge bases, web tables, and relational databases) and services (e.g., web APIs). He has published over 10 papers on question answering/semantic parsing, graph mining and querying, and crowdsourcing at major conferences including SIGKDD, WWW, EMNLP, and CIKM. He has served on the program committee or as reviewer for many top conferences and journals including WWW, ACL, EMNLP, and TKDE. He has interned at Microsoft Research Redmond, IBM T.J. Watson Research Center, and U.S. Army Research Laboratory.