An augmented interpretive framework based on aspect sentiment words aggregation

Given the mounting anxieties surrounding the interpretability of neural models, appraising interpretability remains an unsolved puzzle owing to the ineffectual performance of existing interpretation techniques and evaluation metrics. The architecture of neural network models varies depending on the task at hand, making it challenging to devise a universal method of explanation that can produce coherent justifications for each model. This paper proposes a framework to enhance the interpretability of text sentiment classification models using aspect sentiment words (ASW) aggregation, which can be applied to web services to improve transparency, accountability, and user trust. The proposed method extracts ASW from sentences and consolidates the token importance scores to provide more credible justifications. The paper also introduces new evaluation metrics for faithfulness, which assess whether interpretations accurately reflect the model's decision-making process. The proposed metrics are effective in evaluating the fidelity of rationales to models at the snippet-level.