Year: 2024
Type: article
Abstract: In this EPQ, I built a system that successively enables the large language model to answer questions on the past with evidence from primary sources. This project was motivated by the realization that despite the complexity, the performance of the large language model (LLMs) literally means understanding but simply does pattern recognition, coupled with probabilistic text prediction. Against this backdrop, a specialized AI was developed to help in error correction, given that AI-produced historical data are generally full of errors, by the use of the open-source Langchain-Chatchat that allows integration of LLMs with a structured knowledge database. These contain factual and historical information needed for the AI to compute the accurate answer. The system uses the GLM-4 model by Tsinghua University for the level of GPT-4 and above. This web application is built on Streamlit, an AI-powered system that interacts with the user. The steps through which input queries and historical text were processed by the system include an Unstructured Loader, a Text Splitter, an embedding model, and a Vector Store. Successful implementation of this system not only improved the reliability in AI for historical education but also laid the foundation for further improvement in educational technology and AI interaction models. Future plans will involve further historical database expansion to even more diverse periods and geographies, as well as enhancing the user interface to become more interactive and accessible.
Author: Yuming Zou
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: diamond
OpenAlex ID: W4405389133