Linguistic Features for Interpretable Textual Entailment
SLITE uses 17 linguistic features for textual entailment, reaching 83% on SICK and nearing RoBERTa.
The paper introduces SLITE, a hybrid explainable model for recognizing textual entailment that pairs structural-relational semantic features with distributional measures including entropy and transfer entropy. Logistic regression on 17 features reaches 83% accuracy on three-class SICK and 96% on SICK-CE, outperforming IsoLex by 4 points and landing within 2 points of RoBERTa at much lower cost. Ablation and SHAP analysis show structural-relational features drive most decisions, while distributional features help detect neutrality and contradiction.
- Seventeen features combine entity relations, polarity matching, and embedding alignment.
- Logistic regression scores 83% on SICK and 96% on SICK-CE.
- Within two points of RoBERTa at a fraction of the compute.
- SHAP finds structural features primary and distributional features complementary.
Full article208 words · extracted from arxiv.org · click to collapse
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
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.24932