MultiConIR: Towards Multi-Conditional Information Retrieval

Information retrieval (IR) systems have become central to how we access knowledge in the digital age. However, there's a growing gap between users' complex search needs and the capabilities of current IR systems. While modern users frequently ask multi-faceted questions with multiple constraints ("Find me a thriller movie directed by Christopher Nolan that was released after 2010 and features time travel"), most IR systems and evaluation benchmarks are optimized for simpler, single-condition queries. We proposed MultiConIR, a novel benchmark specifically designed to evaluate retrieval and reranking models under multi-condition retrieval scenarios. Unlike existing datasets that primarily focus on single-condition queries from search engines, MultiConIR captures real-world complexity by incorporating five diverse domains: books, movies, people, medical cases, and legal documents.

From single-conditional retrieval to multi-conditional scenarios.

Overview

MultiConIR (Multi-Condition Information Retrieval) is a comprehensive benchmark aimed at evaluating retrieval models in scenarios involving queries with multiple conditions. Unlike traditional single-condition retrieval tasks, MultiConIR reflects realistic and complex search scenarios across five domains:

• Books
• Movies
• People
• Medical Cases
• Legal Documents

Features

MultiConIR focuses on three key evaluation aspects:

1. Complexity Robustness: How effectively retrieval models handle queries with increasing complexity (1 to 10 conditions).
2. Relevance Monotonicity: Assessing if models maintain consistent relevance ranking when conditions progressively increase.
3. Query Format Sensitivity: Evaluating the stability of retrieval performance across instruction-style and descriptive-style queries.

Dataset Construction

The construction pipeline of MultiConIR datasets.

MultiConIR utilizes a structured and rigorous pipeline for dataset creation:

1. Condition Sentence Extraction: Identifies ten key, non-redundant condition sentences from real-world documents using GPT-4o.
2. Query Generation: Creates instruction-style (structured) and descriptive-style (natural language) queries with incremental conditions from 1 to 10.
3. Hard Negative Generation: Produces semantically similar yet subtly distinct negative sentences to challenge retrieval systems.

Benchmark Tasks

MultiConIR defines three comprehensive evaluation tasks:

• Complexity Robustness: Measures how retrieval performance is affected as the number of conditions in queries increases.
• Relevance Monotonicity: Evaluates models’ ability to rank documents consistently based on the number of conditions matched.
• Query Format Invariance: Assesses model sensitivity to different query formulations (instruction-style vs descriptive-style).

Performance Highlights

• Traditional IR systems deteriorate as query conditions increase, with rerankers suffering more severely.
• Models demonstrate systematic failure in maintaining relevance monotonicity across conditions.
• Reranking models outperform retrievers on simpler queries but experience significant performance drops as query complexity rises.
• Mean pooling focuses on early query conditions while <EOS> pooling biases toward later conditions
• GritLM-7B demonstrates the highest robustness against increased query complexity.
• NV-Embed shows exceptional adaptability to longer documents, maintaining performance stability better than other models.

✅ TODO List

• Release datasets & evaluation code
• Launch MultiConIR-v2: We plan to release the MultiConIR-v2 benchmark, extending the original MultiConIR to fully adapt the MTEB (Massive Text Embedding Benchmark) framework.

Getting Started

Installation

Clone this repository and install dependencies:

git clone https://github.com/EIT-NLP/MultiConIR.git
cd MultiConIR
pip install -r requirements.txt

Data and Models

Datasets and scripts for evaluation are provided in the repository. Refer to the datasets and models folders for further details.

Citation

Please cite our paper if you use MultiConIR in your research:

@misc{lu2025multiconirmulticonditioninformationretrieval,
      title={MultiConIR: Towards multi-condition Information Retrieval}, 
      author={Xuan Lu and Sifan Liu and Bochao Yin and Yongqi Li and Xinghao Chen and Hui Su and Yaohui Jin and Wenjun Zeng and Xiaoyu Shen},
      year={2025},
      eprint={2503.08046},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2503.08046}, 
}
}

Contact

For questions or feedback, please reach out to lux1997@sjtu.edu.cn.