CLMN: Concept based Language Models via Neural Symbolic Reasoning

1. Problem Statement

Deep learning models, particularly in Natural Language Processing (NLP), suffer from a "black box" nature, limiting their deployment in high-stakes domains like healthcare and finance where interpretability is mandatory.

• Limitations of Existing Solutions:
  • Concept Bottleneck Models (CBMs): While successful in Computer Vision, existing CBMs in NLP often enforce rigid binary concept activations. This causes significant information loss in textual representations, leading to a drop in prediction accuracy.
  • Concept Embedding Models (CEMs): These project concepts into continuous spaces to preserve performance but often obscure human-understandable semantics, making the reasoning process opaque.
  • Lack of Dynamic Interaction: Current methods fail to capture dynamic interactions between concepts (e.g., how negation like "not severe" modifies a concept like "disease severity"), relying instead on static, rigid labeling.

2. Methodology: Concept Language Model Network (CLMN)

The authors propose CLMN, a novel neural-symbolic framework that reconciles high performance with transparency by integrating continuous concept embeddings with fuzzy logic-based reasoning.

A. Core Architecture

The framework operates through three main stages:

1. Pretrained Language Model (PLM) Backbone:
   • Input text $x$ is encoded into a latent representation $z$ using models like BERT, RoBERTa, GPT-2, or LSTM.
2. Concept Layer & Embedding:
   • Concept Layer: Explicitly encodes the state probabilities of $S$ concepts into a structured vector. Each concept $s$ has three states: Positive ($p^+$), Negative ($p^-$), and Unknown ($p^u$).
   • Continuous Embeddings: Instead of binary activation, the model learns continuous embeddings for active ($C^+_s$) and inactive ($C^-_s$) states. A scoring function $p_s$ determines the probability of activation, creating a weighted final embedding $\hat{c}_s$.
   • Feature Fusion: These concept embeddings are concatenated with the original text features and passed through a fusion layer to enhance the final prediction, ensuring the model leverages both raw text and concept-aware representations.
3. Neural-Symbolic Reasoning Branch:
   • Polarity Network ($\Phi$): A Multi-Layer Perceptron (MLP) that learns the directional influence (positive or inhibitory) of each concept on the target class.
   • Relevance Network ($\Psi$): A parallel MLP that learns the context-aware importance weight of each concept for a specific instance.
   • Fuzzy Logic Reasoning: The model synthesizes these signals using differentiable fuzzy logic operators (Conjunction $\land$, Disjunction $\lor$, Negation $\neg$) to generate a prediction rule:
     $$\hat{y}_j(x) = \bigwedge_{c_s \in C} (\neg I_{p,s,j} \lor I_{r,s,j})$$
     This allows the model to derive predictions based on logical rules composed of concepts while maintaining differentiability for gradient-based optimization.

B. Training Objective

The total loss function combines three components:
$$L = L_{prediction} + \alpha_1 L_{concept} + \alpha_2 L_{neural}$$

• $L_{prediction}$: Standard cross-entropy for the final task label.
• $L_{concept}$: Loss for predicting concept labels (ensuring the concept layer is accurate).
• $L_{neural}$: Loss for the neural-symbolic reasoning branch (ensuring the logical rules align with the ground truth).

3. Key Contributions

1. Novel Framework (CLMN): The first neural-symbolic framework for NLP that effectively addresses the accuracy-interpreability trade-off. It moves beyond rigid binary concepts to continuous, learnable concept interactions.
2. Neural-Symbolic Reasoning for Interactions: Introduces a mechanism to explicitly model how concepts influence each other and the final prediction using learnable rules and fuzzy logic, solving the issue of static concept relationships.
3. Performance Preservation: Demonstrates that adding concept-aware representations and symbolic reasoning does not significantly degrade task performance compared to standard black-box models.
4. Data Augmentation Strategy: Utilizes LLMs (ChatGPT) to generate new concepts and noisy labels, expanding the concept space beyond human-annotated datasets (CEBaB).

4. Experimental Results

The framework was evaluated on the aug-CEBaB-yelp dataset (restaurant review sentiment analysis) using various backbones (BERT, RoBERTa, GPT-2, LSTM).

• Task Performance:
  • CLMN achieved competitive accuracy with baseline models. For example, with BERT-base, the final task accuracy was 69.26% (vs. 69.49% baseline), a negligible drop of ~0.23%.
  • The Macro F1 score remained high at 79.62%.
• Interpretability Performance:
  • Concept Accuracy: The model achieved high accuracy in predicting intermediate concepts (e.g., 85.85% for BERT), proving it correctly identifies semantic features like "food," "noise," and "price."
  • Reasoning Accuracy: The neural-symbolic reasoning branch achieved 65.35% accuracy, demonstrating that the logical rules derived from concepts can independently predict the final label with reasonable success.
• Ablation Studies:
  • Removing the concept loss weakened the model's ability to learn meaningful concepts.
  • Using only the neural-symbolic component without concept supervision led to performance drops, confirming the need for the dual-loss strategy to align feature learning with logical reasoning.

5. Significance and Impact

• Paradigm Shift: CLMN establishes a new paradigm for NLP by unifying neural representations (for performance) and symbolic reasoning (for interpretability) in a single continuous space.
• Trustworthy AI: By providing transparent derivation pathways (e.g., "The review is negative because the food was good but the noise was loud and the price was high"), the model builds trust in critical sectors like healthcare and finance.
• Dynamic Logic: The use of fuzzy logic allows the model to handle linguistic nuances like negation and context modification, which traditional binary CBMs miss.
• Future Directions: The authors suggest extending this framework to hierarchical concept reasoning and multilingual adaptability, paving the way for more robust and explainable AI systems.

Code Availability: The implementation and data are open-sourced at https://github.com/MichaelYang-lyx/CLMN.