Commonsense Knowledge with Negation: A Resource to Enhance Negation Understanding

Zijie Wang, MohammadHossein Rezaei, Farzana Rashid, Eduardo Blanco

TL;DR

While Large Language Models (LLMs) excel at many tasks, they notoriously stumble when processing negation. This paper introduces a novel, automated pipeline to augment existing commonsense knowledge bases with negation, generating over 2 million high-quality if-then triples. Pre-training open-source LLMs on this data significantly boosts their negation reasoning capabilities across multiple downstream benchmarks without degrading their core performance.

What Problem Does This Paper Address?

Negation appears in roughly 25% of English sentences, yet it remains a major blind spot for state-of-the-art language models. If you tell a model "If a person is hungry, they want to eat food," it understands. But flip it to "If a person is not hungry..." and the model's reasoning often falls apart.

Existing Commonsense Knowledge Bases (CSKBs) like ATOMIC focus almost entirely on affirmative statements. The few attempts to introduce negation have relied heavily on expensive human annotation and only negated the initial condition (the if event), ignoring what happens when you negate the consequence (the then event). This paper aims to bridge the gap between commonsense reasoning and negation understanding at scale without relying on human annotators.

Figure 1: A commonsense knowledge triple with the In- tention relation. Negations are added to both if and then events. Adding different negation cues results in new triples that align (same color on both sides) or conflict with (different colors) commonsense knowledge.

Methods

The authors developed a fully automated framework consisting of three main phases:

  1. Automated Negation Generation: Taking existing if-then triples $\langle A, R, B \rangle$ from ATOMIC and ANION, they used Llama 3.1 70B to syntactically insert the negation cue "not" into the if event, the then event, or both. This yielded three new variations per triple: $\langle \neg A, R, B \rangle$, $\langle A, R, \neg B \rangle$, and $\langle \neg A, R, \neg B \rangle$.
  2. Task-Specific LLM Judge: Automatically generated triples can be Valid (true), Invalid (false), or Ambiguous. Surprisingly, off-the-shelf frontier models like GPT-4o and Claude 3.5 Sonnet struggle to classify these accurately. To solve this, the authors fine-tuned a task-specific judge (Llama 3.1 70B) using QLoRA on a noisy, synthetically generated dataset.
  3. Contrastive Pre-training: Instead of just feeding the model valid facts, they designed a contrastive training objective. The target models (e.g., Llama 3.1 8B, Qwen2 7B) were trained on pairs of Valid and Invalid negated triples to explicitly learn how negating different parts of a sentence flips or preserves its real-world validity.

Datasets

The paper utilizes and introduces several key datasets:

  • Existing CSKBs: ATOMIC (449k affirmative triples) and ANION (142k triples with negated if events).
  • New Corpora: $\neg$ATOMIC (1.79M triples) and $\neg$ANION (285k triples), both fully categorized by their automated judge.
  • Human Benchmark: A newly curated evaluation benchmark of 7,200 triples sampled from ATOMIC's test split, manually annotated by experts to evaluate the LLM judges.
  • Downstream Benchmarks: Evaluated on CondaQA (reading comprehension), NLI with Negation (RTE-Neg, SNLI-Neg, MNLI-Neg), and NevIR (information retrieval).

Key Insights and Ideas

  • Frontier Models Lack Negation Intuition: One of the most striking findings is that zero-shot frontier models like GPT-4o and Claude 3.5 Sonnet are poor judges of negated commonsense. A smaller model explicitly fine-tuned on the task vastly outperforms them.
  • Learning from Mistakes: Training models on Invalid triples is just as important as training them on Valid ones. By comparing contrastive pairs, the model develops a structural understanding of how negation alters logic.
  • No Catastrophic Forgetting: A common fear in fine-tuning is that fixing a specific weakness degrades general performance. Testing on CommonsenseQA proved that pre-training on negated text does not harm the model's ability to reason about standard affirmative concepts.

Key Results

Pre-training on the newly generated corpora yielded substantial, statistically significant improvements across the board:

  • CondaQA: Pre-training improved Llama 3.1 8B's accuracy from 68.7% to 71.5% and Qwen2 7B's accuracy from 65.1% to 69.7%.
  • NLI with Negation: On the RTE-Neg dataset, Llama 3.1 8B saw a massive jump from 60.0% to 81.3% accuracy after training on the new data.
  • Information Retrieval (NevIR): Pairwise accuracy—which requires a model to correctly flip document rankings when a query is negated—improved significantly for Qwen2 7B (from 29.8% to 39.8%) and Llama 3.1 8B (from 30.6% to 42.2%).

Limitations and Future Work

While highly effective, the authors note a few limitations:

  • Scope of Relations: The work is restricted to the if-then relations present in ATOMIC and ANION.
  • Simple Negation Cues: The pipeline only introduces the explicit logical cue "not". It does not explicitly generate morphological negations (like prefixes un- or dis-) or implicit negations (like lack or prevent), though downstream testing showed some positive cross-type transfer.
  • Model Scale: Due to compute constraints, downstream training and evaluation were limited to 7B and 8B parameter models. Testing on larger scales remains future work.

Summary updated

Abstract

Negation is a common and important semantic feature in natural language, yet Large Language Models (LLMs) struggle when negation is involved in natural language understanding tasks. Commonsense knowledge, on the other hand, despite being a well-studied topic, lacks investigations involving negation. In this work, we show that commonsense knowledge with negation is challenging for models to understand. We present a novel approach to automatically augment existing commonsense knowledge corpora with negation, yielding two new corpora containing over 2M triples with if-then relations. In addition, pre-training LLMs on our corpora benefits negation understanding.
BibTeX
@inproceedings{wang2026commonsense,
  author = {Zijie Wang and MohammadHossein Rezaei and Farzana Rashid and Eduardo Blanco},
  title = {{Commonsense Knowledge with Negation: A Resource to Enhance Negation Understanding}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.578},
}

Loading related content...

Figure