Abstract
Recognizing Textual Entailment (RTE) often requires navigating complex linguistic phenomena like negation and quantification. This paper presents a latent discriminative model that integrates the formal framework of Natural Logic with statistical learning. By treating phrase alignments as latent variables and training on sentence-level labels, the model avoids the need for expensive, human-annotated alignment data. Experimental results on a specialized Japanese dataset demonstrate that this weakly supervised approach outperforms models trained on human alignments, suggesting that the model discovers more effective logical paths than those typically intuited by humans.
Introduction
A fundamental challenge in Natural Language Processing is determining the semantic relationship between two sentences—whether one entails, contradicts, or is independent of the other. Traditional approaches often rely on surface-level similarity or word overlap, which struggle with nuances like negation ("I played" vs. "I didn't play") or quantifiers ("Some birds" vs. "All birds").
Natural Logic offers a solution by decomposing sentences into edits and composing their meanings through set-theoretic relations. However, implementing Natural Logic typically requires gold-standard alignments between phrases, which are costly to produce and may not be optimal for computational reasoning. This work addresses this bottleneck by proposing a model that learns these alignments as hidden variables.
The Latent Discriminative Framework
The proposed model unifies statistical learning with Natural Logic through a three-layer architecture. This structure allows the system to move from raw text to a final semantic judgment through a series of principled steps.
- Alignment Layer: This layer identifies potential mappings between bunsetsu (Japanese base phrases) in the premise (T) and the hypothesis (H). It uses an extended MANLI algorithm to explore the space of possible alignments.
- Projection Layer: Contextual factors, such as negation or downward-entailing quantifiers, are accounted for here. These factors "project" or transform the local semantic relations identified in the alignment layer.
- Composition Layer: Finally, the model applies the formal rules of Natural Logic to aggregate these local, projected relations into a single sentence-level prediction (e.g., Equivalence, Forward-Entailment, Contradiction).
Feature Engineering and Factors
To guide the model's latent decisions, the authors employ a wide array of features categorized into specific "factors" (represented by $\Psi$ in the architecture). These factors capture lexical, syntactic, and semantic information across the different layers.
As shown in the table above, the model utilizes features such as WordNet relations, part-of-speech (POS) sequences, and case markers. These features allow the model to weigh different alignment and composition possibilities statistically, even without direct supervision on the alignments themselves.
Weak Supervision: Learning from Sentence Labels
The core innovation of this approach is its use of weak supervision. The model is trained exclusively on sentence-level labels (e.g., "this pair is a contradiction"). During training, the specific alignments and local relations are treated as latent variables. The model iteratively "guesses" which internal configurations best explain the final label and updates its parameters to favor those configurations. This allows the model to find "globally optimal" connections that might differ from human-annotated "gold" alignments but are more effective for the final entailment task.
Evaluation and Key Insights
The researchers evaluated the model on a new Japanese dataset of 598 sentence pairs, specifically designed to test logical reasoning across categories like quantifiers, negation, and modality.
The results highlighted a significant advantage for the latent approach:
- Performance: The latent model achieved 47.5% overall accuracy, outperforming the alignment-supervised model (43.7%).
- Alignment vs. Accuracy: While the supervised model was better at mimicking human alignments, the latent model was superior at the actual entailment task. This suggests that human-chosen alignments can sometimes be a "distractor" for the specific logical composition rules used by the model.
- Logical Strengths: The model proved particularly robust in identifying Contradiction and Forward-Entailment, where structural logic is paramount.
Conclusion
By combining the mathematical rigor of Natural Logic with the flexibility of latent discriminative learning, this work provides a path forward for RTE systems that are both linguistically grounded and data-efficient. While the model currently faces challenges with complex structural transformations (like active-to-passive shifts), its ability to outperform human-aligned systems marks a significant step toward more autonomous and effective semantic reasoning.
Summary updated