duoT5

duoT5

A pairwise reranking model based on T5 that takes a query and two candidate documents as input, predicting which document is more relevant. It extends monoT5 from pointwise to pairwise ranking.

How It Works

1. Input format: "Query: {q} Document0: {d₁} Document1: {d₂} Relevant:"
2. Output: T5 generates a token — the probability of generating “true” for Document0 being more relevant than Document1
3. Aggregation: Pairwise preferences are aggregated across all document pairs to produce a final ranking (e.g., via a sorting-based approach)

Architecture

• Built on T5 (Text-to-Text Transfer Transformer) — a sequence-to-sequence model
• Fine-tuned on MS MARCO or similar relevance datasets
• Typically used as a second-stage reranker after monoT5 or monoBERT narrows down candidates

Multi-Stage Pipeline

Query → BM25 (1000 docs) → monoT5 (rerank to top-50) → duoT5 (pairwise rerank top-50)

Why Pairwise?

Comparing two documents side-by-side lets the model make finer-grained relevance distinctions than scoring each document independently (pointwise). The downside: $O(n^2)$ comparisons, so only feasible on small candidate sets.

Key Properties

• Higher effectiveness than pointwise monoT5 on small candidate sets
• Computationally expensive: $\left(\genfrac{}{}{0ex}{}{n}{2}\right)$ inference calls for $n$ documents
• Practical only as final-stage reranker on top-$k$ (small $k$)

Connections

• Extends monoT5 from pointwise to pairwise
• Part of the Multi-Stage Ranking pipeline
• Related to Cross-Encoder (both are expensive, high-quality rerankers)
• Pairwise approach connects to Learning to Rank (pairwise LTR losses like RankNet)

Appears In

• IR-L05 - Neural IR Intro & Reranking
• PTR (Lin et al.) §3.4.1