L2R: Lifelong Learning for First-Stage Retrieval with Backward-Compatible Representations

By Y. Cai et al

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Task Description
4. Methodology

Summary

This paper introduces L2R, a method for lifelong learning in first-stage retrieval to adapt to evolving data distributions efficiently. It addresses the challenges of learning with unlabeled new documents and avoiding repeated representation computation. The approach involves diverse support negative selection and a ranking alignment objective for backward-compatible representation learning. The method is evaluated on benchmarks and outperforms existing baselines in lifelong learning for retrieval models.