Uncovering relationship between model performance and feedback hyperparameters using explanations and Human-in-the­Loop

Sammanfattning

Warning This paper contains examples of hateful and offensive language. The field of NLP has witnessed substantial growth, revolutionizing sectors from healthcare to social media. However, the advent of complex, black­ box models has led to concerns about their opacity and interpretability. In response, Explainable Artificial Intelligence and Human-In-The-Loop strategies have emerged to address these issues. This paper is part of the Interaction Framework for Artificial and Natural Intelligence (IFAN), a project aimed at enhancing transparency in NLP systems by combining Explainable Artificial Intelligence (XAI) and Human-In-The-Loop (HitL). IFAN provides an interface enabling users to interact with NLP models in real time, offering insights into model behavior through explanations. Additionally, it allows users to modify samples in the training phase through a feedback loop. The study focuses on investigating the relationship between model performance and the feedback hyperparameters. The goals involve fine-tuning BERT­ based models on a hate speech dataset, utilizing HitL and XAI to provide feedback to the model, and exploring feedback hyperparameters to optimize model performance. Two different approaches for utilizing HitL and XAI were explored. First, the existing rationale within the hate speech dataset was utilized as the training dataset for the feedback loop. Second, human annotators were invited to annotate misclassified samples, creating modified data for the model. Despite various hyperparameter configurations, the utilization of the first approach yielded comparatively less improvement on model performance. In conclusion, combining explanations and human-in­ the-loop can improve performance of deep learning models, provided that the right feedback hyperparameters are selected and rationales for creating the training data are thoughtfully considered. In the future, experiments can be conducted to validate the consistency of these relationships across diverse data sets.