Natural Language and Multimodal Text-Tabular Explanations using Large Language Models and SHAP

Abstract

Explainability is a critical pillar of responsible AI, however, there remain several unexplored areas in the field. This thesis focuses on two key areas to enhance interpretability for the end user: utilising large language models (LLMs) to describe, in text, what the data in a table is showing, and creating and applying a new framework to applying SHapley Additive exPlanations (SHAP) explanations to text-tabular datasets. Firstly, this thesis introduces the novel task of taking a set of performance metrics - such as accuracy, precision and F1 score - and fine-tuning LLMs to explain this information. To do so, we collect and provide a dataset and experiment with a deep encoding of the metric information to enable clearer comprehension of the data table. The second chapter extends the text generation approach to explain classification decisions, using a second novel dataset of expert-written explanations to explain a numerical explanation: a set of feature importance values indicating which input the underlying model found most important to the decision. In fine-tuning LLMs on this dataset, experimenting with augmentation and a simplified question-answer task, we demonstrate the capacity to generate understandable and accurate natural language explanations. Further capitalising on the theme of explainability across multiple modalities, this thesis provides a solution to the inability to generate a numerical explanation for text-tabular datasets. Specifically, this thesis proposes a novel multi-modal masker that facilitates the production of SHAP values for any text-tabular dataset, for any method of combining the two modalities. In an extensive analysis, this thesis reveals the issues that arise when adapting the multi-modal dataset to a single modality (text) and applying the existing unimodal masker. Subsequently, we examine the impact that combination strategies and language models have on SHAP values. Finally, we apply the proposed method to a veterinary dataset, using the generated explanations to carry out a deep-dive on which features models found most important and the reasons why PetBERT, an LLM pre-trained on a veterinary corpus, performs better than BERT, a general LLM.