BERT-based questions answering on close domains: Preliminary Report

Natural language processing has seen a revolution in recent years thanks to Large Language Models (LLMs), which are based on generative technologies and set new standards for the field’s main tasks (sentiment analysis, text classification, question answering, etc.). The main issue today with current LLMs are the hallucinations, which cause incomplete control over the model’s entire output and can lead to disastrous outcomes in critical contexts. This makes it impractical to use LLMs in a lot of contexts where a certain level of security and safety is required. We aim to develop a model that can’t hallucinate and reduce false replies, that can be more efficient in terms of time compared to various generative models, and that provides the possibility to explain and identify errors (if any). This is done by avoiding the use of LLMs based on the generation of text and instead using a model that selects the most relevant part of the text and, with an adequate reformulation of the sentence, provides the user with the required pieces of information. We use hotel policies and rules as a case study, but the proposed approach could be applied to all cases that involve questions about a given text. It is important to notice that this work does not require any type of fine-tuning or training on the particular data, making generalisations to other fields and contexts easy.