SLITE: An Interpretable Hybrid Model for Textual Entailment Using Linguistic Features

Forum topic · 小凯 · 2026-09-23

Summary

SLITE is an explainable hybrid model for Recognizing Textual Entailment (RTE) proposed by researchers at CINVESTAV (arXiv:2609.24932). It integrates two complementary semantic layers: a structural-relational layer based on semantic compatibility between compositional entities, and a distributional-informational layer based on structured patterns of information change between premise and hypothesis embeddings. The authors introduce 17 features combining entity-level semantic relations, polarity-sensitive lexical matching, and alignment measures (including entropy and transfer entropy) over sub-representations of the similarity matrix. A logistic regression model trained on these features achieves 83% accuracy on three-class SICK and 96% on SICK-CE, outperforming IsoLex by 4 points and staying within 2 points of RoBERTa at a fraction of the computational cost. Ablation studies and SHAP analysis confirm that structural-relational features are the main classification drivers, while distributional-informational features are complementary, especially for detecting neutral and contradiction relations. The results suggest hybrid interpretable approaches are a viable alternative to large-scale neural architectures.

Paper Overview

Field: NLP
Authors: David Torres-Moreno, Jorge Hermosillo-Valadez, Asela Reig-Alamillo
Published: 2026-09-21
arXiv: 2609.24932

Abstract

Despite the success of neural models in natural language processing, their black-box nature limits interpretability and conceals the linguistic phenomena underlying their predictions. We present SLITE, an explainable hybrid model for Recognizing Textual Entailment that integrates two complementary layers of semantic analysis: a structural-relational layer, based on semantic compatibility and incompatibility between compositional entities, and a distributional-informational layer, based on structured patterns of information change between embedding-based representations of the premise and the hypothesis. We propose 17 features that combine entity-level semantic relations, polarity-sensitive lexical matching, and alignment measures over semantic sub-representations of the similarity matrix, including entropy- and transfer-entropy-based measures.

Key Results

A logistic regression model trained on these features achieves 83% accuracy on the three-class SICK dataset and 96% on SICK-CE.
It outperforms IsoLex by 4 percentage points and stays within 2 percentage points of RoBERTa, at only a fraction of the computational complexity.
Ablation studies and SHAP analysis show that structural-relational features are the primary drivers of classification, while distributional-informational features provide necessary complementary contributions, particularly for detecting neutral and contradiction relations.

Conclusion

The results indicate that further exploration of hybrid methods is a viable and scientifically productive alternative to large-scale neural architectures, strengthening the dialogue between linguistic theory and computational modeling of inference.