Grant: Discovery Grants Program - Individual

Advances in Artificial Intelligence (AI), including Large Language Models (LLMs) such as ChatGPT, allow us to develop highly accurate neural text classifiers. One major disadvantage is their lack of explainability due to their black box nature. In this research program, I propose to develop new methods for neural text classifiers based on LLMs that are explainable, to open their black box architecture, at least partially. Explainability brings with it the other benefits of responsible AI (fairness, lack of bias, trust) because, once the system can explain what the model learnt from data and the decisions/outcomes for each text, the users can trust it and detect and correct bias.     The main objectives of this research program are to develop methods that can explain what the models learned from data and methods to explain the individual decisions for each new test data. I also propose to focus on two case studies: applications to legal text mining and mental health text mining.     The methodology will involve several directions of investigation. One method that I applied in my previous work is to select key phrases collected from the training data to generate explanations (the best candidates being the ones that increase the confidence in the classification). A new method that I propose is to modify the training function of an LLM-based classifier to not only achieve high accuracy in classification, but also to generate coherent and pertinent textual explanations. Another avenue, more challenging, will be based on knowledge graphs, to find facts that can be used as explanations, at architecture-level (what groups of neurons store what information) and at test data level (what fact can justify the classification for that test data). These methods will be applied in zero-shot learning setting (using only the knowledge from the LLM), few-shot learning (using a few training examples to adapt a LLM to the task at hand), and fully supervised learning (fine-tuning on large amounts of labelled training data). For all the methods, the generated explanations will be evaluated quantitatively and qualitatively.     Care will be taken to avoid the traps of LLMs. The generated explanations will be automatically checked to ensure that they do not contain offensive language. Equity, diversity, and inclusion principles will be incorporated into the research from the beginning, since the generated explanations will show any bias in the data or methods. De-biasing methods will be implemented, as well as fact checking, to reduce the risk of wrong generated facts.     The novelty of this work consists in the methodology I propose for generating explanations, as well as in the development of new proof-of-concept applications for the two case studies. This work will contribute to advances in responsible AI systems, leading to their wider application in practice. It will strengthen Canada's position as a world leader in AI.