Multi-Granularity Reasoning for Natural Language Inference

Technical Summary: Multi-Granularity Reasoning for Natural Language Inference

Problem Statement
Natural Language Inference (NLI), or Recognizing Textual Entailment, is a fundamental task requiring the determination of logical relationships (entailment, contradiction, or neutral) between a premise and a hypothesis. While transformer-based pre-trained models have achieved remarkable success, existing approaches predominantly rely on final-layer token representations. The authors argue that these single-layer representations are often insufficient for capturing the complex, hierarchical semantic interactions necessary for effective reasoning. Specifically, fine-grained lexical cues, phrasal compositions, and higher-level contextual semantics tend to become entangled or diluted within a single representation space, limiting the model's ability to discern intricate semantic relationships.

Methodology: Multi-Granularity Reasoning Network (MGRN)
To address these limitations, the paper proposes the Multi-Granularity Reasoning Network (MGRN), a framework designed to mimic the human cognitive process of progressing from shallow lexical matching to deep semantic abstraction. The architecture operates through the following stages:

1. Input Preprocessing and Multi-layer Representation: The input consists of a premise and hypothesis pair processed by a BERT-based encoder. Unlike standard approaches that utilize only the final layer, MGRN extracts hidden states from all $L$ layers of the Transformer stack.
2. Interaction Matrix Construction: For each layer $l$, the model explicitly captures interactions between the two sentences by performing element-wise multiplication (Hadamard product) between the token representations of the premise ($h^{(l,i)}_1$) and the hypothesis ($h^{(l,j)}_2$). This generates an interaction tensor $M^{(l)} \in \mathbb{R}^{n \times m \times d}$, encoding word-by-word interactions specific to that layer's abstraction level.
3. Multi-layer Stacking: The interaction matrices from all layers ($M^{(1)}$ to $M^{(L)}$) are concatenated along a new dimension to form a comprehensive multi-layer interaction representation $M$. This aggregation preserves semantic information ranging from surface features to abstract semantics.
4. DenseNet Feature Extraction: The stacked interaction tensor $M$ is passed through a series of Dense Blocks (Densely Connected Convolutional Networks). This component facilitates efficient gradient propagation and feature reuse by concatenating outputs from all previous layers as inputs to subsequent layers, enabling the extraction of high-level semantic features through multi-step nonlinear mapping.
5. Classification: The final feature vector is fed into a fully connected network followed by a softmax function to predict the three-way classification label.

Key Contributions
The paper outlines three primary contributions:

1. Novel Architecture: The proposal of MGRN, which integrates hierarchical semantic features using a structured interactive tensor to significantly enhance inference accuracy.
2. High-Order Interaction Modeling: The model effectively captures high-order semantic interactions between sentence pairs, extending beyond traditional token-level alignment to incorporate complex compositional and phrasal relationships across multiple abstraction levels.
3. Empirical Validation: Extensive experiments on multiple public benchmarks demonstrate that MGRN consistently outperforms competitive baselines, validating its effectiveness, robustness, and general applicability across diverse NLP tasks.

Experimental Results
The authors evaluated MGRN on ten large-scale public benchmarks, including GLUE tasks (MRPC, QQP, STS-B, MNLI, RTE, QNLI), SNLI, MultiNLI, Sci, SICK, and Twi.

• Performance: When integrated with BERT-base and BERT-large, MGRN achieved average accuracy improvements of 0.8% and 0.7%, respectively, across all datasets. Notably, MGRN outperformed RoBERTa-base by 1.5% and RoBERTa-large by 0.5%.
• Ablation Studies: Removing the interaction matrix or replacing DenseNet with average pooling resulted in clear performance drops, confirming the necessity of fine-grained interaction capture and high-level feature extraction. Using only the last BERT layer also degraded performance, supporting the value of multi-layer fusion.
• Robustness: Under adversarial and perturbation tests (e.g., SwapAnt, NumWord, SwapSyn), MGRN demonstrated superior robustness compared to baselines. It maintained stable performance on transformations disrupting surface forms or syntactic regularity, suggesting its reasoning relies more on semantic interaction patterns than fragile surface features.
• Case Studies: Qualitative analysis showed MGRN successfully handling subtle lexical, gender-related, and numerical differences where models like ESIM, BERT, and SyntaxBERT failed, particularly in cases requiring precise semantic discrimination (e.g., distinguishing "12" from "42").

Significance and Claims
The paper claims that the proposed approach addresses the "brittleness" of modern NLI models by explicitly modeling hierarchical semantic interactions. By simulating the human cognitive progression from surface-level understanding to deep comprehension, MGRN uncovers intricate semantic relationships that are often missed by single-granularity approaches. The authors assert that this multi-granularity reasoning capability not only improves benchmark performance but also enhances the model's robustness against distribution shifts and adversarial attacks, bridging the gap between high accuracy on standard tests and genuine reasoning ability. The work aligns with the direction of robustness-oriented evaluation, demonstrating that integrating semantic information across multiple granularities is a viable path toward more reliable natural language understanding.