Semântica de frames, harmonização terminológica e computação: o uso de frames semânticos como princípio organizador para a harmonização de termos e conceitos e a representação do conhecimento em large language models

Abstract

This PhD dissertation aims to investigate the role of Frame Semantics as an organizing principle for an ERP database, contributing to terminological harmonization, as well as automating this task in a Large Language Model. With digital transformation and the use of Large Language Models (LLMs), such as ChatGPT (OpenAI, 2024), many activities that were previously performed manually are now being automated. In the business context, data processing was once carried out separately to manage a company. Today, we have Enterprise Resource Planning (ERP) software, which integrates all areas of an institution to enable data sharing across all corporate sectors, from Human Resources to Logistics. To make communication more efficient and avoid terminological inconsistencies, the International Organization for Standardization (ISO) developed the ISO 860:2007 recommendation on the harmonization of terms and concepts, which seeks to harmonize terminology within a domain. For this task, ISO relied on the assumptions of Wüster (1968, 1974) and the General Theory of Terminology (Felber, 1979), which do not take contextual, cognitive, and variational aspects into account, nor allow for the harmonization of terms from two or more domains simultaneously. Regarding company terminology, different terms are often used to refer to the same concept in different areas. To address this terminological issue, at a theoretical level, Frame Semantics (Fillmore, 1967, 1976, 1982, 1985) was used to support the mapping of the ERP and its components. For terminological study, Frame-based Terminology (Faber, 2012, 2014, 2015) was utilized to analyze and organize the software’s terminology. With this foundation, a semantic-terminological approach for the harmonization of terms and concepts was proposed. On a practical level, a Knowledge Graph was designed with semantic frames of the ERP, and the harmonization methodology was automated in an LLM by developing different prompt strategies for the model to perform harmonization according to the thesis' proposal. Corpus Linguistics was used as the methodology to compile research corpora, allowing for the extraction of terminology and the selection of terms to be harmonized. After selecting the terms, three terminological sets were analyzed, and manual harmonization was proposed for each group. Subsequently, two language models, ChatGPT-4 and ChatGPT-4 32k, were used to automate the semantic-terminological proposal for harmonizing terms and concepts. The experiments carried out in this study showed that Frame Semantics serves as an organizing principle for ERP and contributes to harmonization, providing more linguistic data for selecting the main term and its variants. Regarding automatic harmonization in an LLM, it was observed that the model at temperature 1 produced harmonization like the one performed manually, as it considered all contextual information from terminological definitions and knowledge representation combined with its creativity. It was observed that a semantic-terminological database enhances the harmonization task by providing more context and linguistic content. Moreover, it was demonstrated that it is possible to automate harmonization in an LLM, achieving results similar to the manual harmonization. Finally, we highlight that the combination between Frame Semantics, Frame-based Terminology and Artificial Intelligence allowed the elaboration of the ERP conceptual structure, as well as it contributed to the organization of the software’s terminology, allowing to give a sturdy design to the harmonization of terms and concepts and to automatize such task in a Large Language Model in a more precise and efficient way.