| تعداد نشریات | 62 |
| تعداد شمارهها | 2,323 |
| تعداد مقالات | 19,157 |
| تعداد مشاهده مقاله | 61,581,639 |
| تعداد دریافت فایل اصل مقاله | 31,230,628 |
Reviewing generative AI models for answering Persian historical questions | ||
| Journal of Digital Content Management | ||
| مقالات آماده انتشار، پذیرفته شده، انتشار آنلاین از تاریخ 31 تیر 1405 | ||
| نوع مقاله: Original Article | ||
| شناسه دیجیتال (DOI): 10.22054/dcm.2026.90071.1293 | ||
| نویسندگان | ||
| Amin Haghdadi* 1؛ Lattifeh Pourmohammadbagher1؛ Mina Safa2 | ||
| 1Faculty of Statistics, Mathematics, and Computer Science, Allameh Tabatabai, Tehran, Iran | ||
| 2Faculty of Perisan Literature and Foreign Languages, Allameh Tabatabai, Tehran, Iran | ||
| چکیده | ||
| Background: In recent years, large language models have emerged as one of the most significant achievements in artificial intelligence, playing a pivotal role in transforming the field of natural language processing. However, in specialized domains such as history, relying solely on the internal knowledge of these models can lead to phenomena such as informational hallucination; meaning the model may generate fluent but incorrect or unsupported responses. Purpose: The present research proposes a retrieval-augmented generation approach to address this challenge for answering Persian historical questions. Method: In this study, sixteen volumes of Persian translations of historical books were used as the dataset. After undergoing preprocessing the texts were prepared to be retrievable and usable for the question-answering process. The Qwen2.5 7b language model was employed as the core engine for generating answers, utilizing information retrieved from historical sources to produce more accurate and well-documented responses. Results: Quantitative evaluation demonstrated that the proposed model showed significant improvement compared to the base Qwen2.5 model. Specifically, the BLEU score increased from 0.48 to 0.90, and in terms of sentence similarity, the proposed model's performance was superior to that of the base model, managing to partially close the gap with the more powerful GPT-4 model. The results indicate that the proposed model not only outperforms its base version but also achieves competitive, near GPT-4 performance on certain metrics. Conclusions: These improvements suggest that combining Qwen2.5 7b with a retrieval-based approach can be an effective strategy for enhancing the accuracy and reliability of language models in Persian applications. | ||
| کلیدواژهها | ||
| Large Language Models؛ Historical question answering؛ Retrieval-Augmented Generation؛ Information Retrieval؛ Natural Language Processing | ||
|
آمار تعداد مشاهده مقاله: 19 |
||