Rationalizing Neural Predictions (1606.04155v2)

Abstract: Prediction without justification has limited applicability. As a remedy, we learn to extract pieces of input text as justifications -- rationales -- that are tailored to be short and coherent, yet sufficient for making the same prediction. Our approach combines two modular components, generator and encoder, which are trained to operate well together. The generator specifies a distribution over text fragments as candidate rationales and these are passed through the encoder for prediction. Rationales are never given during training. Instead, the model is regularized by desiderata for rationales. We evaluate the approach on multi-aspect sentiment analysis against manually annotated test cases. Our approach outperforms attention-based baseline by a significant margin. We also successfully illustrate the method on the question retrieval task.

Rationalizing Neural Predictions

Introduction

The paper "Rationalizing Neural Predictions" by Tao Lei, Regina Barzilay, and Tommi Jaakkola addresses a critical issue in the application of complex neural models by proposing a method to generate justifications—or rationales—for predictions. The authors present an approach where the generation of these rationales is integrated into the learning process itself, without requiring explicit rationale labels during training. This paper combines a generator and an encoder into a unified framework to yield concise, coherent, and sufficient textual rationales.

Methodology

The proposed approach is centered around two modular components: the generator and the encoder. The generator identifies pieces of input text that could serve as rationales, while the encoder uses these rationales to make predictions. The novelty lies in the collaborative training of these components such that the generator is regularized to produce short and coherent rationales that are still predictive.

The generator creates a distribution over potential rationales, which are then evaluated by the encoder. Importantly, no rationale annotations are provided during training. Instead, the model relies on a set of desiderata: rationales must be short, coherent, and sufficient for prediction.

Applications and Results

The approach is evaluated on two NLP tasks: multi-aspect sentiment analysis and similar text retrieval.

1. Multi-aspect Sentiment Analysis: The model is tested on a dataset of beer reviews with ratings for multiple aspects such as appearance, smell, and palate. The authors demonstrate that their model significantly outperforms baselines, including an SVM model and attention-based neural networks. Particularly, the model achieves extraction accuracy of up to 96% for certain aspects, a substantial improvement over the baselines.
2. Similar Text Retrieval: Applied to the AskUbuntu question-answer forum, the model is trained to retrieve similar questions. The evaluation shows the generated rationales to be highly effective, achieving a Mean Average Precision (MAP) close to that of using question titles, which are concise summaries by design. This underscores the reliability of generated rationales in capturing essential information from longer texts.

Encoder and Generator Design

The generator uses a recurrent convolutional neural network (RCNN) to determine a probability distribution for selecting words as part of the rationale. This results in probabilistic tagging of text sequences where the selection is coherent. The encoder, also an RCNN, processes these selected sequences to predict the target variables.

The training objective combines the prediction loss and a regularizer, which penalizes the length and discontinuity of the rationales, ensuring that the resulting rationales are both short and coherent. This joint cost function is minimized using a sampled approximation of the gradient, allowing the model to be trained end-to-end.

Implications and Future Work

This research contributes to the broader field of interpretable AI by demonstrating that neural models can be made more transparent without sacrificing performance. The modularity of the proposed framework allows for its application across various tasks where interpretability is crucial, such as in medical or legal domains. Future work could explore more complex constraints on rationale generation or apply variance reduction techniques to further stabilize the training process.

Additionally, different encoder architectures, such as deep averaging networks or boosting classifiers, could be investigated for their efficacy with this technique. As rationales are inherently flexible, future enhancements could involve tailoring the model to specific domains for optimal performance.

Conclusion

The framework introduced in "Rationalizing Neural Predictions" represents a significant step toward making neural model predictions understandable. By integrating rationale generation into the learning process and employing modular components, the authors demonstrated that it is possible to generate meaningful justifications for neural network decisions in an unsupervised manner. This work lays the groundwork for future advancements in interpretable AI, providing a robust method for generating rationales that other researchers can build upon.