In this research, we propose a framework to generate human-like question-answer pairs with long or factoid answers automatically and, based on them, automatically evaluate the quality of Retrieval-Augmented Generation (RAG). Our framework can also create datasets that assess hallucination levels of Large Language Models (LLMs) by simulating unanswerable questions. We then apply the framework to create a dataset of question-answer (QA) pairs based on more than 1,000 leaflets about the medical and administrative procedures of a hospital. The dataset was evaluated by hospital specialists, who confirmed that more than 50% of the QA pairs are applicable. Finally, we show that our framework can be used to evaluate LLM performance by using Llama-2-13B fine-tuned in Dutch (Vanroy, 2023) with the generated dataset, and show the method’s use in testing models with regard to answering unanswerable and factoid questions appears promising.

The embedding of words and documents in compact, semantically meaningful vector spaces is a crucial part of modern information systems. Deep Learning models are powerful but their hyperparameter selection is often complex and they are expensive to train, and while pre-trained models are available, embeddings trained on general corpora are not necessarily well-suited to domain specific tasks. We propose a novel embedding method which extends random projection by weighting and projecting raw term embeddings orthogonally to an average language vector, thus improving the discriminating power of resulting term embeddings, and build more meaningful document embeddings by assigning appropriate weights to individual terms. We describe how updating the term embeddings online as we process the training data results in an extremely efficient method, in terms of both computational and memory requirements. Our experiments show highly competitive results with various state-of-the-art embedding methods on different tasks, including the standard STS benchmark and a subject prediction task, at a fraction of the computational cost.