Rationalization through Concepts

Diego Antognini, Boi Faltings

TL;DR

This paper introduces ConRAT, a novel neural model that generates "concepts" – interpretable text snippets – to explain predictions, moving beyond single-rationale approaches. ConRAT achieves state-of-the-art performance in sentiment classification tasks, demonstrating that multi-faceted, human-aligned explanations can be learned unsupervised.

What Problem Does This Paper Address?

Neural networks excel at making predictions but often struggle with interpretability, especially in critical applications. While "rationalization" (selecting input features to justify a prediction) is a common explanation method, it often provides a single, potentially incomplete explanation. Humans, however, tend to explain complex decisions by weighing multiple aspects or "concepts." This paper addresses the challenge of creating models that can generate these richer, multi-faceted explanations, mimicking human reasoning, and ensuring these explanations are both accurate and understandable.

Methods

The core of ConRAT is a novel self-explaining architecture composed of three sub-models:

  • Concept Generator: This module identifies potential "concepts" within the input text. It learns to generate binary masks that highlight specific, contiguous text chunks (concepts) of a predefined length.
  • Concept Selector: This module determines which of the generated concepts are actually present and relevant for a given input, outputting a binary score for each concept.
  • Predictor: This module takes the selected concepts and their presence scores to compute the final prediction. It aggregates the contributions of each selected concept linearly, allowing for a multi-faceted explanation.

To ensure the learned concepts are interpretable, ConRAT employs two novel regularizers:

  • Overlap Regularizer: Penalizes redundancy between concepts, encouraging them to be distinct.
  • Diversity Regularizer: Minimizes the cosine similarity between concept representations, promoting diversity.

Additionally, the authors propose optional techniques like knowledge distillation and concept pruning to further boost performance and interpretability.

Datasets

The paper evaluates ConRAT on two datasets:

  • Amazon Electronics Dataset: A single-aspect sentiment classification dataset.
  • Multi-aspect Beer Reviews Dataset: This dataset, commonly used in rationalization research, contains reviews with ratings for multiple aspects (Appearance, Aroma, Palate, Taste, Overall). ConRAT is trained using only the "Overall" aspect label.

Key Insights and Ideas

  • Human-like Explanations: The core idea is that human explanations are often multi-faceted, relying on several concepts. ConRAT successfully models this by learning multiple, distinct concepts rather than a single rationale.
  • Unsupervised Concept Learning: ConRAT learns interpretable concepts without explicit human annotations for these concepts, relying instead on the overall prediction label. This is a significant step towards more scalable explainability.
  • Concept Alignment with Human Judgment: The paper demonstrates that ConRAT's generated concepts align well with human-defined aspects (e.g., Appearance, Aroma) in the beer review dataset, even when trained only on the overall sentiment.
  • Beyond Single Rationales: ConRAT's ability to outperform methods that rely on single rationales highlights the limitations of simpler explanation approaches for complex decisions.

Key Results

  • State-of-the-Art Performance: ConRAT achieves superior predictive and rationale performance compared to existing rationalization methods on both single- and multi-aspect sentiment classification tasks.
  • Improved Rationale Quality: On the Amazon dataset, ConRAT's factual rationales align better with human annotations than baselines, and it shows promise in identifying counterfactual rationales with more concepts.
  • Human-Aligned Concepts: In subjective evaluations on beer reviews, ConRAT's concepts were easily understood by humans, allowing them to accurately guess the aspect and sentiment.
  • Impact of Concept Count: Increasing the number of concepts generally improved performance, up to a point, after which interpretability started to decrease. Pruning concepts to the most dissimilar ones further boosted performance.

Limitations and Future Work

  • Concept Length Constraint: The hard constraint on concept length, while aiding grounding, might limit flexibility. The paper notes that relaxing this constraint led to inferior results.
  • Interpretability vs. Performance Trade-off: While ConRAT aims for interpretable concepts, the paper acknowledges that the diversity desideratum can sometimes be at odds with maximizing prediction accuracy, as spurious correlations might be picked up.
  • Future Work: The authors suggest exploring more sophisticated ways to learn concepts, potentially incorporating external knowledge or more advanced regularization techniques. They also hint at adapting the model for regression problems.

Summary updated

BibTeX
@inproceedings{antognini2021rationalization,
  author = {Diego Antognini and Boi Faltings},
  title = {{Rationalization through Concepts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  url = {https://openreview.net/forum?id=jkNCDX8xY5C},
}
Figure