[Paper Review] Query Performance Prediction using Relevance Judgments Generated by Large Language Models

Chuan Meng, Negar Arabzadeh+3 more|arXiv (Cornell University)|Apr 1, 2024

Data Quality and ManagementDecision Sciences3 citations

TL;DR

This paper proposes QPP-GenRE, a novel query performance prediction framework that decomposes QPP into independent relevance judgments per document using LLMs, enabling accurate prediction of multiple IR metrics and improved interpretability. It achieves state-of-the-art performance on TREC-DL 19–22 by combining LLaMA fine-tuned with human-labeled judgments and an efficient approximation strategy for recall-based metrics.

ABSTRACT

Query performance prediction (QPP) aims to estimate the retrieval quality of a search system for a query without human relevance judgments. Previous QPP methods typically return a single scalar value and do not require the predicted values to approximate a specific information retrieval (IR) evaluation measure, leading to certain drawbacks: (i) a single scalar is insufficient to accurately represent different IR evaluation measures, especially when metrics do not highly correlate, and (ii) a single scalar limits the interpretability of QPP methods because solely using a scalar is insufficient to explain QPP results. To address these issues, we propose a QPP framework using automatically generated relevance judgments (QPP-GenRE), which decomposes QPP into independent subtasks of predicting the relevance of each item in a ranked list to a given query. This allows us to predict any IR evaluation measure using the generated relevance judgments as pseudo-labels. This also allows us to interpret predicted IR evaluation measures, and identify, track and rectify errors in generated relevance judgments to improve QPP quality. We predict an item's relevance by using open-source large language models (LLMs) to ensure scientific reproducibility. We face two main challenges: (i) excessive computational costs of judging an entire corpus for predicting a metric considering recall, and (ii) limited performance in prompting open-source LLMs in a zero-/few-shot manner. To solve the challenges, we devise an approximation strategy to predict an IR measure considering recall and propose to fine-tune open-source LLMs using human-labeled relevance judgments. Experiments on the TREC 2019 to 2022 deep learning tracks and CAsT-19 and 20 datasets show that QPP-GenRE achieves state-of-the-art QPP quality for both lexical and neural rankers.

Motivation & Objective

• Address the limitations of existing QPP methods that rely on single scalar predictions and lack interpretability.
• Overcome the 'one-size-fits-all' issue where a single score fails to represent diverse, non-correlating IR evaluation metrics.
• Enable interpretable QPP by generating per-document relevance judgments that explain prediction outcomes.
• Improve QPP quality by identifying and rectifying errors in generated relevance judgments.
• Ensure scientific reproducibility by using open-source LLMs like LLaMA for relevance judgment generation.

Proposed method

• Decompose QPP into subtasks of predicting relevance for each document in a ranked list using LLMs.
• Use automatically generated relevance judgments as pseudo-labels to predict any IR evaluation measure.
• Implement an approximation strategy to estimate recall-oriented metrics (e.g., RR@10, nDCG@10) by judging only a limited number of top-ranked documents.
• Fine-tune LLaMA via parameter-efficient fine-tuning (PEFT) on human-labeled relevance judgments to improve generation quality.
• Leverage the independence of document-level judgments to enable parallelized inference and scalability.
• Use a judging depth (n) as a hyperparameter to balance accuracy and computational cost in QPP inference.

Figure 1 . Prompt used by LLMs for automatic generation of relevance judgments.

Experimental results

Research questions

• RQ1 Can LLM-generated relevance judgments improve the accuracy and interpretability of QPP across diverse IR evaluation metrics?
• RQ2 How does the approximation strategy for recall-based metrics affect QPP performance while reducing computational cost?
• RQ3 To what extent does fine-tuning LLaMA on human-labeled judgments enhance the quality of relevance judgments for QPP?
• RQ4 Can QPP-GenRE generalize across different rankers (lexical and neural) and IR metrics (precision- and recall-oriented) without retraining?
• RQ5 What is the trade-off between QPP accuracy and inference latency when varying the number of documents judged per query?

Key findings

• QPP-GenRE achieves state-of-the-art performance on TREC-DL 19–22 tracks for both lexical and neural rankers across precision- and recall-oriented metrics.
• The framework significantly outperforms existing QPP baselines such as NQA-QPP, BERTQPP, qppBERT-PL, and M-QPPF in terms of correlation with official IR metrics.
• Using a judging depth of n=10, QPP-GenRE achieves inference latency of 452.60ms per query, which is lower than several supervised baselines when predicting multiple metrics.
• Fine-tuning LLaMA with PEFT on human-labeled judgments improves the quality of generated relevance judgments, directly enhancing QPP accuracy.
• The approximation strategy for recall-based metrics maintains high predictive accuracy while drastically reducing the need to scan the entire corpus.
• QPP-GenRE enables error analysis by translating QPP inaccuracies into observable relevance judgment errors (e.g., false positives/negatives), supporting targeted model improvement.

(a) BM25 on TREC-DL 19