Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Paraphrasing in Affirmative Terms Improves Negation Understanding

MohammadHossein Rezaei, Eduardo Blanco

TL;DR

This paper tackles the challenge of negation understanding in NLP by introducing automatically generated affirmative interpretations—paraphrases that express the same content without negation. It presents two generation methods, T5-HB trained on Large-AFIN (AHB) and T5-CG trained on a ChatGPT paraphrase dataset (ACG), and demonstrates that appending these interpretations to inputs containing negation improves performance across CondaQA and five additional NLU benchmarks. The results are achieved with RoBERTa-Large in a task- and architecture-agnostic setup, without requiring parallel corpora for the affirmative interpretations. The findings indicate that negation robustness can be substantially enhanced through automatically produced paraphrases, with potential extensions to other languages and larger language models.

Abstract

Negation is a common linguistic phenomenon. Yet language models face challenges with negation in many natural language understanding tasks such as question answering and natural language inference. In this paper, we experiment with seamless strategies that incorporate affirmative interpretations (i.e., paraphrases without negation) to make models more robust against negation. Crucially, our affirmative interpretations are obtained automatically. We show improvements with CondaQA, a large corpus requiring reasoning with negation, and five natural language understanding tasks.

Paraphrasing in Affirmative Terms Improves Negation Understanding

TL;DR

This paper tackles the challenge of negation understanding in NLP by introducing automatically generated affirmative interpretations—paraphrases that express the same content without negation. It presents two generation methods, T5-HB trained on Large-AFIN (AHB) and T5-CG trained on a ChatGPT paraphrase dataset (ACG), and demonstrates that appending these interpretations to inputs containing negation improves performance across CondaQA and five additional NLU benchmarks. The results are achieved with RoBERTa-Large in a task- and architecture-agnostic setup, without requiring parallel corpora for the affirmative interpretations. The findings indicate that negation robustness can be substantially enhanced through automatically produced paraphrases, with potential extensions to other languages and larger language models.

Abstract

Negation is a common linguistic phenomenon. Yet language models face challenges with negation in many natural language understanding tasks such as question answering and natural language inference. In this paper, we experiment with seamless strategies that incorporate affirmative interpretations (i.e., paraphrases without negation) to make models more robust against negation. Crucially, our affirmative interpretations are obtained automatically. We show improvements with CondaQA, a large corpus requiring reasoning with negation, and five natural language understanding tasks.
Paper Structure (18 sections, 2 figures, 13 tables)

This paper contains 18 sections, 2 figures, 13 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Generating Affirmative Interpretations
  4. Experimental Results
  5. CondaQA
  6. Qualitative and Error Analysis
  7. Other NLU Tasks
  8. Conclusion
  9. Future Work.
  10. Paraphrases vs. Affirmative Interpretations
  11. NLU Corpora
  12. Attempting to Generate Affirmative Interpretations with ChatGPT
  13. Affirmative Interpretations Examples
  14. Training Details with CondaQA
  15. Training Details with Additional NLU Tasks
  16. ...and 3 more sections

Figures (2)

  • Figure 1: Attempting to generate affirmative interpretations with ChatGPT results in a nonsensical conversation. ChatGPT appears to be able to identify negations yet uses them when instructed to not do so
  • Figure 2: An example from CondaQA. The negation in the original sentence is lack. The crowdworkers wrote a paraphrase of the original sentence, which is included in the edited passage ([…] by the absence of mobile charge carriers). The question is written based on the original paragraph and answered based on the original and all three edited passages (only paraphrase edit shown). The answer to the question (for the edited passage) is Yes. The dataset does not explicitly indicate the edited sentence. However, we extract it as explained in Appendix \ref{['sec:condaqaexample']}.