SEA: September - Generative IR

Location: L3.36, Lab 42, Science Park Amsterdam

The Zoom link, in case you want to join online, will be available when you RSVP 1 day prior to the event.

Speaker #1: Gustavo Penha (Spotify)
Title: Generative Retrieval for Search and Recommendation
Abstract: In this talk, I will go over research conducted at Spotify on generative models for the retrieval step in search and the candidate generation step for recommendation. I will talk about the construction of item IDs, i.e., how we represent items as tokens in the generative model, for music search and investigations on a joint search and recommendation model including the creation of semantic IDs that work in a joint S&R setting.
Bio: Gustavo Penha is a research scientist at Spotify. He has a PhD from TU Delft in conversational search and recommendation. His research interests include information retrieval, recommender systems, natural language processing, and machine learning

Speaker #2: Zhaochun Ren (Universiteit Leiden)
Title: Scaling Instruction-Finetuning for Zero-Shot Generative Retrieval
Abstract: In this talk, I will introduce our recent study on InstructGR, a zero-shot generative retrieval (GR) framework. GR reformulates information retrieval as generating document identifiers, offering end-to-end optimization and natural integration with language models. However, GR often struggles to generalize to unseen tasks. InstructGR addresses this by leveraging natural language instructions and introducing: (i) an LM-based docid generator for heterogeneous documents, (ii) an instruction-tuned query generator for diverse corpus indexing, and (iii) a reverse annealing decoding strategy balancing precision and recall. I will also briefly discuss our broader efforts to enhance the generalization of GR.
Bio: Dr. Zhaochun Ren is an Associate Professor at Leiden University, the Netherlands. He is interested in information retrieval and natural language processing, with an emphasis on conversational AI, recommender systems, and information retrieval. He aims to develop intelligent agents that can address complex user requests and solve core challenges in NLP and IR towards that goal. His research has been recognized with multiple awards at RecSys, SIGIR, WSDM, EMNLP, and CIKM. Prior to joining Leiden, he was a Professor at Shandong University and a Research Scientist at JD.com.

Counter: SEA Talks #289 and #290.