Rationale-Augmented Convolutional Neural Networks for Text Classification.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Text classification is a fundamental task in Natural Language Processing.
• Existing methods often overlook the explicit rationale behind document labels.
• Supervised learning for text classification typically uses document-level labels.

Purpose of the Study:

• To develop a Convolutional Neural Network (CNN) model for text classification that integrates both document and sentence-level supervision.
• To effectively utilize explicit sentence rationales provided by annotators to enhance classification performance.
• To create a model that provides interpretable predictions by highlighting supporting sentences.

Main Methods:

• A hierarchical Convolutional Neural Network (CNN) architecture is proposed.
• The model represents documents as a linear combination of sentence representations.
• A sentence-level convolutional model estimates the probability of a sentence being a rationale.
• Sentence contributions to the document representation are scaled based on rationale probability.

Main Results:

• The proposed model consistently outperforms strong baselines across five diverse classification datasets.
• The approach demonstrates superior performance when leveraging both document labels and associated rationales.
• The model naturally generates explanations for its classification decisions.

Conclusions:

• Jointly exploiting document and sentence-level rationales significantly improves text classification accuracy.
• The hierarchical CNN model offers an effective way to integrate multi-granularity supervision.
• The model's ability to provide explanations enhances its utility and trustworthiness in real-world applications.