Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ScoNe: Benchmarking Negation Reasoning in Language Models With Fine-Tuning and In-Context Learning

Jingyuan Selena She, Christopher Potts, Samuel R. Bowman, Atticus Geiger

TL;DR

ScoNe introduces a controlled benchmark to analyze negation semantics and its scope in natural language inference, incorporating ScoNe-NLI and ScoNe-NLG. The study shows that RoBERTa/DeBERTa achieve near-perfect ScoNe-NLI performance with substantial fine-tuning, while in-context learning with InstructGPT largely fails to capture negation scope in NLI tasks. In contrast, InstructGPT demonstrates strong performance on ScoNe-NLG narrative completion, indicating the model can reason about negation when the task aligns with its pretraining. The authors propose an interpretability framework based on causal abstraction and simple algorithms to diagnose whether models implement explicit negation-scope reasoning. Collectively, ScoNe provides a precise diagnostic tool for probing semantic negation and informs prompt design and future interpretability research.

Abstract

A number of recent benchmarks seek to assess how well models handle natural language negation. However, these benchmarks lack the controlled example paradigms that would allow us to infer whether a model had learned how negation morphemes semantically scope. To fill these analytical gaps, we present the Scoped Negation NLI (ScoNe-NLI) benchmark, which contains contrast sets of six examples with up to two negations where either zero, one, or both negative morphemes affect the NLI label. We use ScoNe-NLI to assess fine-tuning and in-context learning strategies. We find that RoBERTa and DeBERTa models solve ScoNe-NLI after many shot fine-tuning. For in-context learning, we test InstructGPT models and find that most prompt strategies are not successful, including those using step-by-step reasoning. To better understand this result, we extend ScoNe with ScoNe-NLG, a sentence completion test set that embeds negation reasoning in short narratives. Here, InstructGPT is successful, which reveals the model can correctly reason about negation, but struggles to do so on prompt-adapted NLI examples outside of its core pretraining regime.

ScoNe: Benchmarking Negation Reasoning in Language Models With Fine-Tuning and In-Context Learning

TL;DR

ScoNe introduces a controlled benchmark to analyze negation semantics and its scope in natural language inference, incorporating ScoNe-NLI and ScoNe-NLG. The study shows that RoBERTa/DeBERTa achieve near-perfect ScoNe-NLI performance with substantial fine-tuning, while in-context learning with InstructGPT largely fails to capture negation scope in NLI tasks. In contrast, InstructGPT demonstrates strong performance on ScoNe-NLG narrative completion, indicating the model can reason about negation when the task aligns with its pretraining. The authors propose an interpretability framework based on causal abstraction and simple algorithms to diagnose whether models implement explicit negation-scope reasoning. Collectively, ScoNe provides a precise diagnostic tool for probing semantic negation and informs prompt design and future interpretability research.

Abstract

A number of recent benchmarks seek to assess how well models handle natural language negation. However, these benchmarks lack the controlled example paradigms that would allow us to infer whether a model had learned how negation morphemes semantically scope. To fill these analytical gaps, we present the Scoped Negation NLI (ScoNe-NLI) benchmark, which contains contrast sets of six examples with up to two negations where either zero, one, or both negative morphemes affect the NLI label. We use ScoNe-NLI to assess fine-tuning and in-context learning strategies. We find that RoBERTa and DeBERTa models solve ScoNe-NLI after many shot fine-tuning. For in-context learning, we test InstructGPT models and find that most prompt strategies are not successful, including those using step-by-step reasoning. To better understand this result, we extend ScoNe with ScoNe-NLG, a sentence completion test set that embeds negation reasoning in short narratives. Here, InstructGPT is successful, which reveals the model can correctly reason about negation, but struggles to do so on prompt-adapted NLI examples outside of its core pretraining regime.
Paper Structure (39 sections, 1 figure, 7 tables)

This paper contains 39 sections, 1 figure, 7 tables.

Table of Contents

  1. Introduction
  2. A Brief Review of Negation in NLI Benchmarks
  3. ScoNe-NLI
  4. Fine-Tuning on ScoNe-NLI
  5. In-context Learning on ScoNe-NLI
  6. InstructGPT makes systematic errors similar to a baseline that ignores negation entirely.
  7. ScoNe-NLG
  8. Dataset
  9. In-context Learning on ScoNe-NLG
  10. Future Work on Interpretability
  11. Conclusion
  12. Experimental Details
  13. Fine-tuning Protocol
  14. Hugging Face Models
  15. Fine-Tuning Datasets
  16. ...and 24 more sections

Figures (1)

  • Figure 1: Four human-interpretable algorithms for ScoNe-NLI. The first two solve the task perfectly, and the other two implement flawed heuristics that a model might learn to implement. The function get-lexrel retrieves the relation between the aligned words in the premise and hypothesis, count-scoped counts scoped negations, count-neg counts negations regardless of scope, and get-first returns true if the first negation scopes, while get-second returns true if there is a second negation and it scopes.