A1671
Title: Enhancing embedding models through specialized fine-tuning in the banking sector
Authors: Claudia Berloco - Intesa Sanpaolo (Italy) [presenting]
Enrico Capuano - Politecnico di Torino (Italy)
Abstract: The widespread development of generative AI and natural language processing has led to the adoption of embedding models in various applications, including question-answering systems. These systems rely on the representation of words and sentences through embeddings to retrieve relevant information before feeding the large language models (LLMs). However, open-source pre-trained multipurpose embedding models may not capture specific nuances in certain contexts, such as the banking sector. Fine-tuning pre-trained embedding models on a dedicated dataset is investigated to improve their performance in specific contexts. A proprietary dataset is constructed from banking sector documents and divided into training and test sets. Various pre-trained open-source multipurpose embedding models are evaluated on the test set and fine-tuned on the training set. Performance is also assessed within a retrieval augmented generation (RAG) pipeline. The results are compared to those of the original multipurpose models to determine the impact of fine-tuning on sentence comprehension and retrieval. Analysis of the fine-tuned models' performance provides insight into tailoring embedding models to meet the unique needs of various industries and applications.
