Neural Information Retrieval with Weak Supervision

In recent years, machine learning approaches, and in particular deep neural networks, have yielded significant improvements on several natural language processing and computer vision tasks; however, such breakthroughs have not yet been observed in the area of information retrieval (IR). Besides the complexity of the IR tasks, such as understanding the user's information needs, a main reason is the lack of large-scale training data for many IR tasks. This necessitates studying how to design and train machine learning algorithms where there is no large-scale data in hand. In this talk, I will introduce training neural networks for IR tasks with weak supervision, where labels are obtained automatically without human annotations (e.g., click data). I will also present the application of such learning strategy in relevance ranking, word embedding, and query performance prediction.
 

Hamed Zamani is a fourth-year PhD candidate (with distinction) in the Center for Intelligent Information Retrieval at the University of Massachusetts Amherst, working with W. Bruce Croft. His research interests include various aspects of core information retrieval, such as query representation, document representation, and ranking. His research mostly focuses on unsupervised or weakly supervised approaches. Hamed is an active member of the IR community. He is an ACM SIGIR student liaison, representing North and South Americas. He has served as a PC member at major IR venues, such as SIGIR, WSDM, WWW, RecSys, and CIKM, and has co-organized two workshops at the SIGIR and RecSys conferences in 2018.