Designing and Evaluating Context-Sensitive Visualization Models for Deep Learning Text Classifiers

As Transformer models gain complexity in the realm of Natural Language Processing (NLP), their interpretability becomes a significant challenge. To tackle this issue, visual explanation emerges as a compelling avenue. Central to visual explanation is the ability to visualize the model’s path leading to specific outputs, shedding light on the relevant features or components that influence the desired outcomes. One major objective of NLP visual explanation is to emphasize the most salient portions of the text that exert the most significant impact on the model’s output. Numerous visual explanation techniques for NLP models have surfaced in recent times. However, evaluating and comparing the performance of these methods presents a major hurdle. Conventional classification accuracy measures are inadequate for assessing visualization quality. To address this, rigorous criteria are essential to gauge the usefulness of the extracted insights for explaining the models. Additionally, visualizing discrepancies in the knowledge extracted by different models becomes crucial for effective ranking purposes. This is an area of research with very few available results. In this work, we investigate how to to evaluate explanations/visualizations resulting from Machine Learning (ML) models for text classification. We describe and apply several methods for evaluating the quality of text visualizations, including both automated techniques based on quantifiable measures and subjective techniques based on human judgements.