پرهام رحيمي

With the rapid growth of unstructured texts on the web, relation extraction plays a vital role in natural language processing research, automatically transforming unstructured texts into structured formats and aiding in completing and updating existing knowledge graphs with up-to-date data. Relation extraction methods typically require a substantial amount of labelled training data, which can be time-consuming and costly. Distant supervision has been proposed to address this challenge. Although traditional relation extraction models have achieved significant success, recent advancements in large language models have demonstrated remarkable achievements in advancing knowledge graphs in various natural language processing tasks, making them worthy of investigation in this task. This thesis outlines the use of pro‎mp‎t engineering and large language models to overcome existing challenges in improving state-of-the-art results in relation extraction. To this end, a method for generating engineered pro‎mp‎ts based on traditional relation extraction datasets is presented. Subsequently, datasets of 24 and 56 relations, each containing over 60,000 and more than 600,000 pro‎mp‎ts, are provided for relation extraction with distant supervision approach based on the New York Times corpus, for large language models. Finally, the T5 language model is fine-tuned to provide better results (1.12 increase in F-score and 2.44 increase in Precision) than current state-of-the-art models. Furthermore, the use of few-shot and zero-shot learning with the large language model LaMa2 is recommended to achieve better results in relation extraction for less frequent relations in the training dataset.

استخراج رابطه , نظارت از راه دور , مدل‌هاي زباني , مدل‌هاي زباني بزرگ , پردازش زبان طبيعي

Relation Extraction , Distant Supervision , Language Models , Large Language Models , Natural Language Processing

Parham Rahimi

Behrouz Minaei-Bidgoli