NevIR: Negation in Neural Information Retrieval

TL;DR

This paper investigates how negation affects neural information retrieval models, revealing that most models perform poorly on negation cases and highlighting the need for better negation understanding in IR systems.

Contribution

The paper introduces a benchmark for evaluating negation handling in neural IR models and analyzes performance across different architectures, exposing significant gaps in current models' capabilities.

Findings

Cross-encoders perform best on negation tasks

Most IR models perform at or below random chance on negation

Fine-tuning and larger models improve performance but do not close the gap

Abstract

Negation is a common everyday phenomena and has been a consistent area of weakness for language models (LMs). Although the Information Retrieval (IR) community has adopted LMs as the backbone of modern IR architectures, there has been little to no research in understanding how negation impacts neural IR. We therefore construct a straightforward benchmark on this theme: asking IR models to rank two documents that differ only by negation. We show that the results vary widely according to the type of IR architecture: cross-encoders perform best, followed by late-interaction models, and in last place are bi-encoder and sparse neural architectures. We find that most information retrieval models (including SOTA ones) do not consider negation, performing the same or worse than a random ranking. We show that although the obvious approach of continued fine-tuning on a dataset of contrastive documents containing negations increases performance (as does model size), there is still a large gap between machine and human performance.