Transparency and accountability in AI decision support: Explaining and visualizing convolutional neural networks for text information

摘要

Proliferating applications of deep learning, along with the prevalence of large-scale text datasets, have revolutionized the natural language processing (NLP) field, thereby driving the recent explosive growth. Nevertheless, it is argued that state-of-the-art studies focus excessively on producing quantitative performances superior to existing models, by playing “the Kaggle game.” Hence, the field requires more effort in solving new problems and proposing novel approaches and architectures. We claim that one of the promising and constructive efforts would be to design transparent and accountable artificial intelligence (AI) systems for text analytics. By doing so, we can enhance the applicability and problem-solving capacity of the system for real-world decision support. It is widely accepted that deep learning models demonstrate remarkable performances compared to existing algorithms. However, they are often criticized for being less interpretable, i.e., the “black box.” In such cases, users tend to hesitate to utilize them for decision-making, especially in crucial tasks. Such complexity obstructs transparency and accountability of the overall system, potentially debilitating the deployment of decision support systems powered by AI. Furthermore, recent regulations are emphasizing fairness and transparency in algorithms to a greater extent, turning explanations more compulsory than voluntary. Thus, to enhance the transparency and accountability of the decision support system and preserve the capacity to model complex text data at the same time, we propose the Explaining and Visualizing Convolutional neural networks for Text information (EVCT) framework. By adopting and ameliorating cutting-edge methods in NLP and image processing, the EVCT framework provides a human-interpretable solution to the problem of text classification while minimizing information loss. Experimental results with large-scale, real-world datasets show that EVCT performs comparably to benchmark models, including widely used deep learning models. In addition, we provide instances of human-interpretable and relevant visualized explanations obtained from applying EVCT to the dataset and possible applications for real-world decision support.