Explaining generative model for long-form question answering with reasoning graph

Question Answering (QA) is a branch of Natural Language Processing (NLP) that focuses on constructing systems that can answer human inquiries in natural language. There are various approaches to achieve this objective, the primary distinction is whether the system is open or closed-domain, and in the case of open-domain, whether it is text-based, knowledge-based, or a hybrid. Even if the model can store domain information in its parameters and leverage it to answer questions, open-domain QA systems fail to encode all the external knowledge needed, such as Wikipedia text corpus, because of the vast amount of it. That would be theoretically possible for Large Language models with billions of parameters. This approach has two important drawbacks: (i) training such model requires highly expansive hardware, enormous datasets, and a lot of time; (ii) updating the knowledge is a not-trivial task that involves retrain. Thus, it is better to leverage external Knowledge Bases (EKB) such as Knowledge Graph (KG) that models world knowledge via relational triplets between entities. This approach is called retrieval-based QA. Despite the high accuracy of these models, the reasoning behind the generated answer is not transparent, indeed these models are called black box. This has raised the question of whether we can trust these answers. This is now a major area of research known as Explainability. This work examines how to improve open-domain QA responses by introducing a custom layer that interacts with a knowledge graph.