Scholarlink

<p>The development of Arabic conversational agents using deep learning techniques, like Pre-Trained Language Models (PTLMs) and Generative Artificial Intelligence (AI), is an exciting challenge in the realm of Natural Language Processing. Researchers have extensively used deep learning techniques to develop conversation agents in languages other than Arabic, but they need to make more progress in Arabic language studies. Current NLP techniques for Arabic language dialogue systems mainly rely on manually crafted feature-based and rule-based approaches. Although these methods can be effective, they also have limits when it comes to dealing with the complexities of Arabic morphology, different ways of writing words, and the potential for ambiguous meanings. The integration of recent successes in Generative AI and Pre-Trained Language Models offers a promising approach to address these limitations and improve the performance of Arabic dialogue systems. This paper proposes a novel Neural Network architecture based on an Encoder-Attention-Decoder (EAD-Bi-LSTM) architecture that incorporates Bidirectional Long-Short-Term Memory (Bi-LSTM) on both the encoder and decoder, along with attention mechanisms. This architecture aims to effectively capture discriminative features and learn essential word weights that significantly impact disease detection in Arabic dialogue systems. The approach employs feature embedding based on two BERT variants, AraBERT and ARBERT, to generate comprehensive, semantic, and syntactic word representation vectors. This approach requires minimal external knowledge or hand-crafted feature engineering, allowing the model to generate rich-word representations autonomously. The Encoder-Attention-Decoder-ARBERT model demonstrated strong effectiveness and adaptability across various domains compared to existing work in different metrics such as accuracy, precision, recall, F1 score, C@1, perplexity and blue score. Our evaluation highlighted the generalizability of the model, with promising results observed in multiple sectors, and particularly notable success in medical applications.</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

EAD-Bi-LSTM-BERT: a novel deep learning architecture for arabic question answering systems

  • 期刊论文

  • 发表日期：

    2025年7月19日

查看全文
  • Yassine Saoudi,
  • Mohamed Mohsen Gammoudi

摘要

The development of Arabic conversational agents using deep learning techniques, like Pre-Trained Language Models (PTLMs) and Generative Artificial Intelligence (AI), is an exciting challenge in the realm of Natural Language Processing. Researchers have extensively used deep learning techniques to develop conversation agents in languages other than Arabic, but they need to make more progress in Arabic language studies. Current NLP techniques for Arabic language dialogue systems mainly rely on manually crafted feature-based and rule-based approaches. Although these methods can be effective, they also have limits when it comes to dealing with the complexities of Arabic morphology, different ways of writing words, and the potential for ambiguous meanings. The integration of recent successes in Generative AI and Pre-Trained Language Models offers a promising approach to address these limitations and improve the performance of Arabic dialogue systems. This paper proposes a novel Neural Network architecture based on an Encoder-Attention-Decoder (EAD-Bi-LSTM) architecture that incorporates Bidirectional Long-Short-Term Memory (Bi-LSTM) on both the encoder and decoder, along with attention mechanisms. This architecture aims to effectively capture discriminative features and learn essential word weights that significantly impact disease detection in Arabic dialogue systems. The approach employs feature embedding based on two BERT variants, AraBERT and ARBERT, to generate comprehensive, semantic, and syntactic word representation vectors. This approach requires minimal external knowledge or hand-crafted feature engineering, allowing the model to generate rich-word representations autonomously. The Encoder-Attention-Decoder-ARBERT model demonstrated strong effectiveness and adaptability across various domains compared to existing work in different metrics such as accuracy, precision, recall, F1 score, C@1, perplexity and blue score. Our evaluation highlighted the generalizability of the model, with promising results observed in multiple sectors, and particularly notable success in medical applications.