Commonsense Knowledge with Negation: A Resource to Enhance Negation Understanding

Abstract: Negation is a common and important semantic feature in natural language, yet 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.

• The paper introduces an automated system that generates negated if-then commonsense triples to enhance LLM sensitivity to negation.
• It employs LLM-based validation to accurately label over 2 million negation-augmented triples, improving outcomes in QA, NLI, and IR tasks.
• Pre-training with these negated resources yields significant accuracy gains on negation-sensitive tasks without impairing affirmative knowledge.

Commonsense Knowledge with Negation: Enhancing Negation Understanding in LLMs

Motivation and Problem Statement

Negation is a pervasive semantic phenomenon, present in approximately a quarter of English sentences, yet remains a persistent failure mode in LLMs during natural language understanding tasks. Despite substantial progress in commonsense knowledge base (CSKB) construction, there has been minimal focus on representing and reasoning with negated commonsense assertions. Prior resources such as Atomic focus on positive instances, while the only direct effort, Anion, handles negation solely on the antecedent side through manual annotation, resulting in limited scale and scope.

This work introduces an automated, scalable methodology to systematically augment existing if-then commonsense corpora with negations—on both conditions and consequences—enabling the exploration of a more comprehensive set of negated commonsense triples. The resulting resource comprises over 2 million triples and demonstrates its efficacy for improving negation sensitivity in LLMs across tasks requiring nuanced handling of negation.

Figure 1: A knowledge triple with the Intention relation, demonstrating how negating the condition (if event), consequence (then event), or both generates new triples that align or conflict with established commonsense knowledge.

Methodology for Negation Augmentation

The authors propose an automated procedure that applies negation to the events within commonsense if-then triples. Each original triple ⟨A, R, B⟩, with A as the condition and B as the consequence, is expanded into up to three new forms by introducing logical negation ("not") at different positions: $\langle \neg A, R, B \rangle$, $\langle A, R, \neg B \rangle$, and $\langle \neg A, R, \neg B \rangle$. This approach substantially generalizes prior methodologies by considering negation on both sides of the implication, and it is fully automated via LLM prompting with syntactic validation.

A major technical challenge is that blindly adding negations does not guarantee the resulting triple remains valid or meaningful. Therefore, the paper introduces an LLM-based automatic judge—fine-tuned and supervised to validate whether any given triple with negation is (i) Valid (aligned with commonsense), (ii) Invalid (contradictory to commonsense), or (iii) Ambiguous (context-dependent or lacking clear causal structure). Training labels are sourced from original corpora, composition strategies, and LLM-synthesized invalid examples.

Manual annotation and inter-annotator agreement establish benchmark datasets, and the pipeline delivers over 2 million validated negated triples in two new corpora: $\neg$Atomic and $\neg$Anion.

Empirical Evaluation: Validation and Downstream Task Transfer

LLMs including Llama 3.1 (8B, 70B), GPT-4o, and Claude Sonnet are benchmarked as both few-shot validators and as fine-tuned instance-level judges. Fine-tuning Llama 3.1 70B achieves superior precision for both valid and invalid negated triples (0.70/0.79). These high-precision labels are shown to be critical for the downstream utility of the resource.

Performance transfer is assessed via pre-training on the newly constructed corpora, with evaluation across five diverse benchmarks covering question answering (CondaQA), natural language inference (RTE-Neg, SNLI-Neg, MNLI-Neg), and information retrieval (NevIR). Across three model classes (RoBERTa-large, Llama 3.1, Qwen2), pre-training on the resource yields consistent, statistically significant improvements on all negation-sensitive tasks compared to baseline and existing unnegated CSKB pre-training.

For example:

• RoBERTa-large pre-trained on the new corpora achieves +5.5% accuracy improvement on CondaQA over the off-the-shelf baseline, outperforming a larger UnifiedQA-v2-large model.
• Open-source LLMs such as Llama 3.1 8B display a +4.1% accuracy improvement on CondaQA after pre-training.
• NevIR pairwise accuracy rises from 24.5% to 34.3% with RoBERTa-large, demonstrating transfer to information retrieval domains requiring negation-aware ranking.

Remarkably, there is no observed degradation on affirmative, non-negated tasks, indicating that the introduced resource does not induce catastrophic forgetting or impaired generalization for positive knowledge.

Ablation and Error Analysis

Ablation studies reveal that pre-training with individual negation types (negating only the condition or consequence) contributes less than leveraging the full spectrum of negated forms. Furthermore, data quality and scale are critical: training with randomly labeled or insufficiently sized datasets leads to significant performance drops, especially for complex IR settings such as NevIR.

Error analyses indicate the resource substantially improves handling of all principal negation phenomena, especially explicit verbal negation, affixal negation, and multiple negation interaction patterns. For example, pre-training improves Llama 3.1 8B's CondaQA verbal negation accuracy by 6.5 points.

Theoretical and Practical Implications

The systematic inclusion of negation in commonsense resources closes a glaring gap in CSKB modeling. The paper demonstrates that negation-augmented corpora are necessary and sufficient for boosting model robustness to negation, establishing a new standard for CSKB completeness. The approach illustrates that explicit negative knowledge and conflict patterns can be reliably generated and validated at scale—an enabling step for downstream interpretability, robust QA, and negation-sensitive dialogue and IR systems.

Practically, this resource enables more accurate LLM deployment in domains where negation is critical (medical, legal, IR) and underlines the insufficiency of raw data scaling or model size alone for robust negation sensitivity.

Future Directions

The framework is currently restricted to binary logical negation and if-then relations. Immediate extensions include covering alternative relation types, incorporating other negation cues (semantic or morphosyntactic), and generalizing beyond the "not" operator. There are clear avenues to extend automated validation and augmentation to richer event representations and multimodal knowledge bases. Model-centric future work could combine the resource with adversarial and contrastive pre-training or leverage the patterns for symbolic reasoning modules.

Conclusion

Systematically enriching commonsense knowledge with negation through automated augmentation and LLM-based validation addresses a central limitation in current LLM-based reasoning. Evaluation across multiple architectural classes and tasks demonstrates measurable, statistically significant improvements in negation understanding with no cost to affirmative knowledge. This work provides a large-scale, high-precision resource and methodology for advancing negation-aware AI systems, setting a foundation for enhanced model reliability in nuanced reasoning scenarios.