Knowledge-enhanced neural models for question answering based on retrieval

Explainability in AI models has emerged as a paramount concern in various domains, including natural language processing (NLP). Understanding and interpreting AI models' decision-making processes is crucial for ensuring their ethical and trustworthy deployment. This thesis addresses the pressing need to improve explainability in language models, explicitly focusing on knowledge retrieval and integration for question-answering tasks. It delves into the rich landscape of neuro-symbolic and sub-symbolic techniques in reasoning, highlighting their strengths in combining rule-based interpretability with data-driven learning. The study provides a comprehensive overview of research advances in explainable AI. Also, it proposes to adopt a context-augmentation strategy for tackling the question-answering task in the biomedical field. This approach aims at enhancing the performances of retrieval models without intervening directly on its core functioning, namely modifying its parameters, but rather carrying out a reranking of the retrieved passages internally to the inference network. The proposed strategy suggests leveraging external sources more efficiently by integrating structured knowledge into the answer generation. This thesis encourages the usage of systems fostering explainability by showing the theoretical foundations and results of knowledge-enhanced solutions in the question-answering field. By inspiring confidence in users, regulators, and stakeholders, we propel the deployment of AI technologies towards a more transparent and accountable future.