Researchers are exploring the integration of large language models (LLMs) with knowledge graphs and ontologies to enhance their performance in specialized financial tasks. Despite the general linguistic capabilities of pre-trained models like GPT, they often fall short in niche fields such as finance. This study suggests a novel approach by combining LLMs with ontological reasoning, using an Enterprise Knowledge Graph (EKG) to infuse structured financial knowledge into the fine-tuning process. The methodology involves generating a comprehensive “reasoning space” using the Vadalog engine, which transforms logical inferences into natural-language statements for training the LLM.
The study highlights several contributions, including the creation of a synthetic, knowledge-rich corpus for LLM fine-tuning aimed at improving performance in tasks like question answering, explanation, translation, and description within the financial domain. Implemented with Google’s T5-large model in a trading environment, the ontology-augmented model demonstrated improved accuracy and transparency over a baseline model.
However, the paper notes limitations such as the need for high-quality domain knowledge and the challenges of maintaining up-to-date ontologies. The computational intensity of ontological reasoning and the potential bias towards the ontology’s content in the LLM are also concerns. Future research should focus on rigorous evaluation, scalability, and maintaining the model’s general language abilities.
Despite these challenges, the approach represents a significant step towards more reliable AI systems in finance, aligning with industry trends that value knowledgeable and interpretable AI. The integration of symbolic reasoning with data-driven learning could reduce the “black-box” nature of LLMs, making them more adept at explaining market behavior.
