TL;DR
Large Language Models (LLMs) struggle with Natural Language Inference (NLI) when premises are not declarative statements (e.g., questions, commands). This paper demonstrates that standard NLI training on datasets like MNLI fails to teach models this crucial distinction, leading to misclassifications. They introduce methods and datasets to improve LLM understanding of these "non-assertive" premises, showing promising results without significantly degrading performance on standard NLI tasks.
What Problem Does This Paper Address?
Natural Language Inference (NLI) is a core task in Natural Language Processing (NLP) where a model determines the relationship between two sentences: a premise and a hypothesis. Traditionally, this relationship is understood through logical inference, where a premise must be a declarative statement that can be true or false. However, real-world language includes questions, commands, and sentence fragments, which, by their nature, cannot be true or false and therefore cannot logically entail or contradict other statements.
The problem this paper addresses is that current state-of-the-art NLI models, particularly those fine-tuned on large datasets like MNLI (Multi-Genre Natural Language Inference), fail to recognize this fundamental distinction. They often incorrectly infer relationships from non-assertive premises, treating questions or commands as if they were factual claims. This oversight can lead to significant errors in applications where logical reasoning is critical, such as legal analysis or educational feedback systems. The authors highlight a tension between the stated goals of NLI datasets and the actual examples they contain, which often include non-assertive premises.
Methods
The authors employ a multi-pronged approach:
- Theoretical Framework: They establish clear criteria (C1 and C2) for a sentence to express a "determinate claim" capable of logical inference. C1 states that only sentences with assertive pragmatic force can express determinate claims, while C2 requires a precisely identified subject and predicate.
- Semantic Parsing for Identification: They utilize the Graphical Knowledge Representation (GKR) parser to automatically identify non-assertive premises (questions, commands, or incomprehensible fragments) within the MNLI dataset. GKR's ability to categorize sentence types is crucial here.
- Dataset Creation:
- Evaluation Datasets: They create a "GKR-n" dataset specifically for evaluating models on non-assertive premises, ensuring all premises are indeed non-assertive and thus should ideally lead to a "neutral" classification. They also use the standard MNLI-Matched evaluation set.
- Fine-Tuning Datasets: They construct eight different fine-tuning datasets. These datasets are designed to teach LLMs to correctly classify pairs with non-assertive premises as "neutral." They systematically vary the proportion of assertive vs. non-assertive premises in entailment/contradiction pairs and the proportion of non-assertive premises in neutral pairs.
- LLM Fine-Tuning and Evaluation: They fine-tune several transformer-based LLMs (DeBERTa-base, XLNET-base, RoBERTa-large) that were already pre-trained on MNLI. They also include a RoBERTa-large model not pre-trained on MNLI to test their assumption that pre-training on MNLI is necessary. Models are evaluated on both the GKR-n and MNLI-M datasets.
- ChatGPT Exploration: They conduct a manual probe of ChatGPT with non-assertive premise pairs to assess its understanding of logical relationships involving such utterances.
Datasets
- MNLI (Multi-Genre Natural Language Inference): Used as the primary source for premise-hypothesis pairs and for initial model fine-tuning.
- GKR-n (GKR-neutral): A newly created evaluation dataset consisting of 406 premise-hypothesis pairs with non-assertive premises, manually verified to be correctly classified as neutral.
- MNLI-Matched (MNLI-M): The standard matched evaluation dataset for MNLI, used to assess performance on general NLI tasks.
- Fine-Tuning Datasets: Eight custom datasets were created, combining unfiltered MNLI samples with GKR-filtered samples (either assertive or non-assertive premises) to teach models about non-assertive premises.
Key Insights and Ideas
- The "Non-Assertive Premise" Problem: The core insight is that standard NLI models trained on datasets like MNLI are fundamentally flawed because they don't distinguish between declarative statements and other utterance types (questions, commands). This leads to incorrect inferences.
- GKR as a Crucial Tool: The GKR parser is vital for identifying non-assertive premises, enabling the creation of targeted datasets for training and evaluation.
- Targeted Fine-Tuning Works: Fine-tuning LLMs on datasets specifically designed to highlight non-assertive premises significantly improves their ability to correctly classify these cases as "neutral."
- Minimal Performance Degradation: The improved understanding of non-assertive premises does not come at a significant cost to performance on standard NLI tasks (like MNLI-M).
- ChatGPT's Nuance: While ChatGPT shows some ability to recognize incomplete premises, it still struggles with commands and questions, indicating that even advanced conversational models aren't fully robust in this area.
Key Results
- MNLI Dataset Issue: The authors estimate that approximately 15% of MNLI premises are non-assertive, meaning about 6.6% of MNLI is potentially mislabeled if these pairs are assigned entailment or contradiction labels.
- GKR Parser Performance: The GKR parser achieved a precision of 83.6% and recall of 63.3% for identifying non-assertive premises, significantly outperforming a simple heuristic based on punctuation.
- Fine-Tuning Effectiveness: Models fine-tuned on the custom datasets showed a dramatic improvement in accuracy on the GKR-n evaluation dataset, reaching up to 89% for RoBERTa-large-mnli, compared to near-random accuracy (32%) for models fine-tuned only on the unfiltered MNLI dataset.
- MNLI-M Performance: Performance on the MNLI-M dataset decreased slightly (e.g., from 87% to 77% for RoBERTa-large-mnli), but the authors argue this is a manageable trade-off, representing a 3.3% effective accuracy loss due to correctly identifying mislabeled non-assertive pairs.
- ChatGPT Performance: ChatGPT correctly identified "neutral" relationships in 44% of the tested non-assertive premise pairs. It showed a better understanding of fragmentary premises than questions or commands.
Limitations and Future Work
- GKR Parser Quality: The fine-tuning datasets' effectiveness is dependent on the GKR parser's accuracy. While good, it's not perfect, meaning some training data might be faulty, potentially limiting further performance gains.
- ChatGPT Sample Size: The exploration of ChatGPT was limited by the manual effort required, suggesting the need for larger-scale studies to confirm findings.
- Future Work: The authors plan to explore more sophisticated fine-tuning methods and larger datasets. They also aim to further probe ChatGPT with larger datasets and investigate its understanding of logical relationships more deeply.
Summary updated