Negation-Aware Neural Information Retrieval

Improving neural information retrieval systems' understanding of negation through targeted pre-training.

Further Pre-training RoBERTa for Negation in Neural Information Retrieval

Negation can substantially change the meaning of a sentence, yet neural information retrieval systems often struggle to account for it when ranking documents.

In this project, we investigate whether targeted further pre-training can improve RoBERTa’s ability to handle negation in neural information retrieval.

We evaluate RoBERTa-large on NevIR, a benchmark designed to test whether retrieval models correctly rank documents when relevance depends on negation.

We study several negation-focused pre-training strategies and evaluate them in two settings:

  • No-supervision: models are evaluated on NevIR without fine-tuning on its training set
  • Fine-tuning: models are further fine-tuned on the NevIR training set

Key Results

  • Negation-focused further pre-training consistently improves RoBERTa-large in both evaluation settings.
  • Combining negation-enhanced Atomic and ANION pre-training produces a 9.1-point improvement in NevIR score in the no-supervision setting compared with the STSB-trained RoBERTa-large baseline.
  • Further pre-training with NSP followed by NevIR fine-tuning achieves a NevIR score of 80.8.
  • This exceeds the reported RankGPT o3-mini score of 77.3 in our comparison.
  • Error analysis shows that targeted pre-training helps correct cases where the model fails to properly account for negation when ranking documents.

Overall, our results suggest that targeted linguistic pre-training can improve neural retrieval models’ sensitivity to negation without requiring additional inference-time computation.

Full Report · Code