Hybrid linguistic features improve explainable textual entailment accuracy
Linguistic Features for Interpretable Textual Entailment
Computation and LanguageSymbolic Computation
Summary
Understanding whether one sentence logically follows from another can be tricky, especially for computers. The authors created a method that uses two types of language analysis: one that looks at how parts of sentences relate to each other, and another that studies how information changes between sentences. This method, called SLITE, uses 17 specific features and can explain why it makes decisions while keeping accuracy high. It performs nearly as well as large AI models but is simpler and more interpretable. This approach helps connect how people understand language with how machines process it.
What this means in practice
- •For natural language processing engineers: Build interpretable systems that decide if one text logically follows from another using linguistic features instead of only large neural models.
- •For legal document analysts: Improve automated tools to flag contradictions or neutrality between texts in contracts by combining linguistic structure and information change measures.
Authors
David Torres-Moreno, Jorge Hermosillo-Valadez, Asela Reig-Alamillo
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 measures based on entropy and transfer entropy. A logistic regression trained on these features achieves an accuracy of 83% on three-class SICK and 96% on SICK-CE, outperforming IsoLex by 4 percentage points and falling within 2 percentage points of RoBERTa with a fraction of its computational complexity. Ablation studies and SHAP analysis confirm that structural-relational features are the primary drivers of classification, while distributional-informational features provide essential complementary contributions, particularly for detecting neutrality and contradiction. Our results demonstrate that further exploration of hybrid approaches is a viable and scientifically productive alternative to massive neural architectures, and we hope they will strengthen the dialogue between linguistic theory and computational modeling of inference