Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthetic Test Collections for Retrieval Evaluation

Hossein A. Rahmani, Nick Craswell, Emine Yilmaz, Bhaskar Mitra, Daniel Campos

TL;DR

The paper tackles the challenge of IR evaluation data scarcity by proposing a fully synthetic approach that generates both queries and relevance judgments with LLMs. It introduces a pipeline that samples MS MARCO passages, filters low-quality seeds, generates queries with a T5 BeIR model and GPT-4, and labels documents via GPT-4 prompts, validating the approach on TREC DL 2023. The results show synthetic collections can reproduce system rankings similar to real judgments, with Kendall's tau around 0.8568 and comparable agreement metrics, although some bias and overestimation tendencies are observed. The work demonstrates the feasibility of scalable, fully synthetic IR evaluation datasets and highlights important directions for bias mitigation and cross-LLM validation.

Abstract

Test collections play a vital role in evaluation of information retrieval (IR) systems. Obtaining a diverse set of user queries for test collection construction can be challenging, and acquiring relevance judgments, which indicate the appropriateness of retrieved documents to a query, is often costly and resource-intensive. Generating synthetic datasets using Large Language Models (LLMs) has recently gained significant attention in various applications. In IR, while previous work exploited the capabilities of LLMs to generate synthetic queries or documents to augment training data and improve the performance of ranking models, using LLMs for constructing synthetic test collections is relatively unexplored. Previous studies demonstrate that LLMs have the potential to generate synthetic relevance judgments for use in the evaluation of IR systems. In this paper, we comprehensively investigate whether it is possible to use LLMs to construct fully synthetic test collections by generating not only synthetic judgments but also synthetic queries. In particular, we analyse whether it is possible to construct reliable synthetic test collections and the potential risks of bias such test collections may exhibit towards LLM-based models. Our experiments indicate that using LLMs it is possible to construct synthetic test collections that can reliably be used for retrieval evaluation.

Synthetic Test Collections for Retrieval Evaluation

TL;DR

The paper tackles the challenge of IR evaluation data scarcity by proposing a fully synthetic approach that generates both queries and relevance judgments with LLMs. It introduces a pipeline that samples MS MARCO passages, filters low-quality seeds, generates queries with a T5 BeIR model and GPT-4, and labels documents via GPT-4 prompts, validating the approach on TREC DL 2023. The results show synthetic collections can reproduce system rankings similar to real judgments, with Kendall's tau around 0.8568 and comparable agreement metrics, although some bias and overestimation tendencies are observed. The work demonstrates the feasibility of scalable, fully synthetic IR evaluation datasets and highlights important directions for bias mitigation and cross-LLM validation.

Abstract

Test collections play a vital role in evaluation of information retrieval (IR) systems. Obtaining a diverse set of user queries for test collection construction can be challenging, and acquiring relevance judgments, which indicate the appropriateness of retrieved documents to a query, is often costly and resource-intensive. Generating synthetic datasets using Large Language Models (LLMs) has recently gained significant attention in various applications. In IR, while previous work exploited the capabilities of LLMs to generate synthetic queries or documents to augment training data and improve the performance of ranking models, using LLMs for constructing synthetic test collections is relatively unexplored. Previous studies demonstrate that LLMs have the potential to generate synthetic relevance judgments for use in the evaluation of IR systems. In this paper, we comprehensively investigate whether it is possible to use LLMs to construct fully synthetic test collections by generating not only synthetic judgments but also synthetic queries. In particular, we analyse whether it is possible to construct reliable synthetic test collections and the potential risks of bias such test collections may exhibit towards LLM-based models. Our experiments indicate that using LLMs it is possible to construct synthetic test collections that can reliably be used for retrieval evaluation.
Paper Structure (6 sections, 3 figures, 3 tables)

This paper contains 6 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Synthetic Test Collection
  3. Synthetic Query Generation
  4. Synthetic Relevance Judgment Generation
  5. Analysis of Bias in System Evaluation
  6. Conclusion and Future Work

Figures (3)

  • Figure 1: Scatter plot of the effectiveness (i.e., NDCG@10) of TREC DL 2023 runs according to the real queries and synthetic queries with human judgments. A point represents a single run averaged over all queries.
  • Figure 2: Scatter plots of the effectiveness (i.e., NDCG@10) of TREC DL 2023 run according to the real queries with human judgments and our synthetic queries with (a) sparse judgments and (b) synthetic judgments.
  • Figure 3: Scatter plots of the effectiveness of TREC DL 2023 runs based on synthetic vs. real test collections to analyse the bias towards systems using the same language model as the one used in synthetic test collection construction.