Scholarlink

<p>Semantic similarity is a fundamental task in natural language processing (NLP), essential for applications such as information retrieval, paraphrase detection, and text clustering. These challenges in Arabic have inspired the development of deep learning-based models that have demonstrated to be performing remarkably well on the limited existing resources. In this paper, a Siamese neural network framework is proposed and evaluated for semantic similarity of Arabic sentences, using pretrained embeddings and different deep learning architectures such as CNN, GRU, LSTM, and BiLSTM. The model is trained and tested on a large dataset collected from five Arabic datasets, including Arabic Paraphrasing Benchmark, Arabic Semantic Similarity dataset, and three datasets of SemEval-2017 Task1. The results demonstrate that the Siamese model based on LSTM and GRU architecture benefit more from AraBERT embeddings while the Siamese model with CNN and BiLSTM improves more with AraGPT embeddings. This refers to the method by which each embedding preserves contextual information, which might be local or sequential. The best performance of the Siamese model utilizing AraBERT is achieved by the LSTM architecture with the lowest error values and the best accuracy value of 0.74, and F1 score of 0.81. While the Siamese CNN-based architecture with AraGPT outperform other networks, maintaining lower error rates and providing the best effective and consistent performance with accuracy of 0.77, F1 score of 0.84, and recall of 0.91. These findings highlight the effectiveness of domain-specific neural network architectures for Arabic Semantic Similarity task and offer guidelines for selecting optimal architectures based on targeted evaluation criteria.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating Semantic Similarity of Arabic Sentences Using Siamese Neural Networks and Deep Learning Models

  • 期刊论文

  • 发表日期：

    2025年11月5日

查看全文
  • Marwah Alian,
  • Hanen Himdi,
  • Khaled Shaalan

摘要

Semantic similarity is a fundamental task in natural language processing (NLP), essential for applications such as information retrieval, paraphrase detection, and text clustering. These challenges in Arabic have inspired the development of deep learning-based models that have demonstrated to be performing remarkably well on the limited existing resources. In this paper, a Siamese neural network framework is proposed and evaluated for semantic similarity of Arabic sentences, using pretrained embeddings and different deep learning architectures such as CNN, GRU, LSTM, and BiLSTM. The model is trained and tested on a large dataset collected from five Arabic datasets, including Arabic Paraphrasing Benchmark, Arabic Semantic Similarity dataset, and three datasets of SemEval-2017 Task1. The results demonstrate that the Siamese model based on LSTM and GRU architecture benefit more from AraBERT embeddings while the Siamese model with CNN and BiLSTM improves more with AraGPT embeddings. This refers to the method by which each embedding preserves contextual information, which might be local or sequential. The best performance of the Siamese model utilizing AraBERT is achieved by the LSTM architecture with the lowest error values and the best accuracy value of 0.74, and F1 score of 0.81. While the Siamese CNN-based architecture with AraGPT outperform other networks, maintaining lower error rates and providing the best effective and consistent performance with accuracy of 0.77, F1 score of 0.84, and recall of 0.91. These findings highlight the effectiveness of domain-specific neural network architectures for Arabic Semantic Similarity task and offer guidelines for selecting optimal architectures based on targeted evaluation criteria.