SEA: Generative IR

In this edition of SEA we will discuss generative information retrieval. We have two amazing speakers lined up: Ruqing Zhang (University of Amsterdam) and Gabriel Bénédict (RTL NL).

This will be a hybrid event, the in-person event will take place at Lab42, Science Park, room L3.36.
***
IMPORTANT: You will be able to view the Zoom link once you 'attend' the meetup on this page.
***
17.00: Ruqing Zhang (University of Amsterdam)
Title: Corpusbrain: Generative Retrieval Models for Knowledge-Intensive Language Tasks
Abstract: Information retrieval is a core task in many real-world applications, such as digital libraries, expert finding, Web search, and so on. Traditional information retrieval systems typically follow the paradigm of “index-retrieve-then-rank” to balance efficiency and effectiveness. Recently, Metzler et al. envisioned a fundamentally different paradigm called model-based IR. The key idea is to fully parameterize different components of index, retrieval and ranking with a single consolidated model. In essence, a generative scheme is adopted to directly predict the relevant document identifiers (docids) with a given query. In this talk, I will first describe several key advantages and characteristics of generative retrieval and then introduce our research works related to Knowledge-Intensive Language Tasks (KILT): Firstly, the off-the-shelf pre-trained generative models are directly adopted to retrieve relevant documents; Secondly, a pre-trained generative retrieval model is designed to improve a variety of downstream search tasks; Thirdly, a unified generative retriever is presented to perform different retrieval tasks at different levels of granularity. The last part of the talk will introduce the industrial application of our method and provide a summary of several existing challenges.
***
***
17.30: Gabriel Bénédict (RTL NL)
Title: RecFusion: A Binomial Diffusion Process for 1D Data for Recommendation
Abstract: Generative Information Retrieval (a.k.a. Generative Neural Search or chatGPT + attribution + no-hallucination) has experienced substantial growth across multiple research communities and has been highly visible in the popular press. Theoretical, empirical, and actual user-facing products have been released that retrieve documents (via generation) (Generative Document Retrieval) or directly generate answers given an input request (Grounded Answer Generation).
A subfield of Generative IR, Generative Recommendations, is still in its infancy. We propose RecFusion to use diffusion models to generate recommendations. We benchmark classical diffusion formulations (normal distribution for the forward and backward diffusion process, Unets and ELBO) against formulations fitted to the RecSys setting: 1D diffusion (user-by-user), binomial diffusion and multinomial loss (like in MultVAE). We also experiment with diffusion guidance to condition the generation of recommendation strips on movie genre (a.k.a. controllable recommendation).

Just keep counting: SEA talks #244 and #245.