MorphNLI: A Stepwise Approach to Natural Language Inference Using Text Morphing

1. Problem Statement

Natural Language Inference (NLI) is the task of determining the logical relationship between a premise ($P$) and a hypothesis ($H$), classifying them as Entailment, Contradiction, or Neutral. While Neural NLI models have achieved high performance, they suffer from several critical limitations:

• Lack of Logical Robustness: They often fail to capture granular logical features such as monotonicity, negation, and quantifier scope.
• Overfitting to Artifacts: Models tend to exploit dataset-specific biases (e.g., assuming coreference between entities in premise and hypothesis) rather than learning true logical reasoning.
• Poor Cross-Domain Generalization: Performance drops significantly when moving from training domains to unseen domains.
• Explainability Issues: Large Language Models (LLMs) often produce "unfaithful" explanations or fail to articulate why a decision was made, sometimes hallucinating reasoning that aligns with training artifacts rather than the input text.

The authors propose MorphNLI, a modular, step-by-step approach designed to decompose the complex NLI decision into a sequence of simpler, explainable transformations.

2. Methodology: MorphNLI

MorphNLI operates on the principle of text morphing, incrementally transforming the premise into the hypothesis through a series of atomic edits. The pipeline consists of four main stages:

A. Voice Normalization (VN)

To handle syntactic rigidity, the premise and hypothesis are first normalized to the active voice using a smaller language model. This prevents the morphing model from "hallucinating" intermediate steps due to passive/active voice shifts.

B. Text Morphing (Stepwise Transformation)

Instead of comparing $P$ and $H$ directly, the system generates a finite sequence of sentences $M = \{M_0, M_1, ..., M_n\}$, where $M_0 = P$ and $M_n = H$.

• Morphism Operations: The transformation uses three atomic operations (similar to Levenshtein distance but at word/phrase granularity):
  1. Replace: $(replace, <old\_text>, <new\_text>)$
  2. Remove: $(remove, )$
  3. Insert: $(insert, )$
• Model Training:
  • Teacher Model: An LLM (GPT-4) is used with In-Context Learning (ICL) to generate synthetic training data. It is prompted with Chain-of-Thought (CoT) to output operations before generating intermediate sentences.
  • Filtering: Generated data is filtered to remove "lazy morphisms" (no intermediate steps) and "short morphisms" (intermediate sentences shorter than $P$ or $H$, which risk losing context).
  • Student Model: A smaller LLM (GPT-4o-mini) is fine-tuned on this filtered synthetic data to perform morphing during inference.

C. Modular NLI Inference

An off-the-shelf NLI classifier (e.g., BART or RoBERTa) is applied to every adjacent pair in the sequence $(M_{i-1}, M_i)$ to generate an intermediate NLI label.

D. Aggregation

The sequence of intermediate labels is aggregated into a final prediction using a simple heuristic:

• If all intermediate labels are Entailment, the final label is Entailment.
• Otherwise, the final label is the first non-entailment label encountered in the sequence (e.g., if the sequence is [Entailment, Entailment, Neutral, Entailment], the result is Neutral).

3. Key Contributions

1. MorphNLI Framework: A novel, modular NLI approach that combines text morphing with neural NLI without requiring additional task-specific supervision (the morphing model is trained on synthetic data).
2. Cross-Domain Superiority: The method demonstrates robust performance in Out-of-Domain (OOD) settings, outperforming state-of-the-art baselines significantly.
3. Explainability: The method produces a "reasoning chain" (the sequence of edits and intermediate labels) that naturally explains the final decision, offering higher faithfulness than LLM-generated explanations.
4. Empirical Validation on LLMs: The study shows that even massive LLMs (GPT-4o) benefit from morphing on logic-heavy datasets, suggesting they lack inherent logical semantics.

4. Experimental Results

The authors evaluated MorphNLI on two datasets: MNLI (Multi-Genre, complex natural language) and SICK (Sentences Involving Compositional Knowledge, logic-heavy).

• In-Domain (ID) Performance: MorphNLI showed a slight drop in ID accuracy compared to "Vanilla" (direct) NLI models. The authors attribute this to the fact that standard NLI models are not trained on incremental transformations.
• Out-of-Domain (OOD) Performance: MorphNLI significantly outperformed baselines in all cross-domain scenarios (e.g., training on MNLI, testing on SICK, and vice versa).
  • Improvements: Relative improvements of up to 12.6% were observed.
  • Specifics: On SICK (OOD), MorphNLI improved RoBERTa accuracy by 1.74% and BART by ~1.74%. On MNLI (OOD), improvements were even larger (5-6%).
• LLM Integration: When GPT-4o was used as the NLI engine, morphing improved its accuracy on SICK but degraded it on MNLI. The authors hypothesize this is due to data contamination: GPT-4o likely overfits to SNLI artifacts (coreference assumptions) present in its training data, which are prevalent in MNLI but absent in SICK.
• Explainability Analysis:
  • Human evaluators rated MorphNLI's explanations as superior to GPT-4o and Llama 3.1 on the SICK dataset, despite MorphNLI using much smaller models.
  • GPT-4o explanations were often incorrect on SICK due to overfitting to SNLI artifacts (incorrectly assuming coreference), whereas MorphNLI's step-by-step edits exposed these logical gaps.
  • LIME analysis confirmed that morphing helps NLI models focus on semantically relevant words rather than spurious correlations.

5. Significance and Conclusion

MorphNLI addresses the "black box" nature of modern NLI by enforcing a logical, stepwise reasoning process. Its significance lies in:

• Robustness: It reduces overfitting to dataset artifacts by breaking complex inference into atomic, verifiable steps.
• Interpretability: It provides a transparent chain of reasoning (the morphing path) that can be audited, unlike the opaque reasoning of standard LLMs.
• Generalization: It proves that decomposing logical tasks improves performance in cross-domain settings, a critical requirement for real-world applications.

The authors conclude that while the approach is currently limited to English and relies on specific LLMs for morphing generation, it offers a promising direction for building more faithful and logically sound NLP systems. Future work could involve fine-tuning NLI models specifically for incremental transformations and extending the method to other languages and reasoning tasks.