RankT5: Fine-Tuning T5 for Text Ranking with Ranking Losses (SIGIR 2023)

Forum topic · 小凯 · 2026-07-05

Summary

RankT5, published at SIGIR 2023, proposes fine-tuning the T5 encoder-decoder language model for text ranking using dedicated ranking losses instead of the usual generative training objectives. The work addresses how to adapt large pretrained sequence-to-sequence models for both reranking (pointwise scoring of query-document pairs) and document retrieval in a multi-task setup. By introducing ranking-specific loss functions tailored to T5's encoder-decoder architecture, RankT5 achieves strong performance on standard information retrieval benchmarks such as MS MARCO, complementing prior listwise approaches like RankBERT and duot5 based on encoder-only or cross-encoder models. The paper is part of a broader research line on applying large language models to learning-to-rank in web search, offering reproducible findings for neural ranking pipelines. This forum entry indexes the paper under the Ranking for Search category, with the original source available at https://dl.acm.org/doi/abs/10.1145/3539618.3592047, and cross-references related work on deep learning to rank, reranking models, and LLM-based ranking.

RankT5: Fine-Tuning T5 for Text Ranking with Ranking Losses (SIGIR 2023)

Overview

RankT5 is a SIGIR 2023 paper that fine-tunes the T5 text-to-text transfer transformer for text ranking tasks using dedicated ranking losses, rather than relying solely on T5's original generative training objective.

Paper: RankT5: Fine-Tuning T5 for Text Ranking with Ranking Losses
Venue: SIGIR 2023
DOI / Source: https://dl.acm.org/doi/abs/10.1145/3539618.3592047
Category: Ranking for Search

Key Idea

Most prior work adapts pretrained language models for ranking in one of two ways:

1. RankBERT-style: fine-tune encoder-only models (e.g., BERT) as pointwise rerankers with ranking losses. 2. duoT5-style: cast ranking as a generative task, asking T5 to predict relevance tokens for pairs of documents.

RankT5 takes a different route: it fine-tunes the full encoder-decoder T5 model directly with ranking-specific losses, enabling it to serve both as a reranker and as a document retrieval model in a multi-task setup. The authors show this ranking-loss-based adaptation of T5 achieves strong performance on standard benchmarks (e.g., MS MARCO), demonstrating that encoder-decoder LMs can be effectively specialized for ranking beyond purely generative formulations.

Why It Matters

Shows that loss design matters when adapting large pretrained models to ranking: ranking losses can be applied to encoder-decoder architectures, not just encoders.
Provides a multi-task framework covering both retrieval and reranking stages of a search pipeline.
Contributes to the broader line of research on LLMs for learning-to-rank, which underpins modern neural search systems and retrieval-augmented generation (RAG).

Related Entries

Deep Learning to Rank in Industrial Search Engines
A Generative Re-ranking Model for List-level Multi-objective Optimization
A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE
Accelerating Listwise Reranking: Reproducing and Enhancing FIRST (SIGIR)
Adaptive Neural Ranking Framework: Toward Maximized Business Goal

> Note: Quantitative results should be verified against the original PDF; this entry summarizes the paper based on its abstract and public metadata.

References

Original paper: https://dl.acm.org/doi/abs/10.1145/3539618.3592047

RankT5: Fine-Tuning T5 for Text Ranking with Ranking Losses (SIGIR 2023)

Summary

RankT5: Fine-Tuning T5 for Text Ranking with Ranking Losses (SIGIR 2023)

Overview

Key Idea

Why It Matters

Related Entries

References

Tags