ارائه ­دهنده:
 آقای وسام المعموری

استاد راهمنا

Dr. Behrouz Minaei-Bidgoli


هیات داوران:
استاد راهنما:  Dr. Behrouz Minaei-Bidgoli
استاد مشاور: Dr. Sayyed Sauleh Eetemadi
Dr.Farzaneh Ghayour Baghbani,
Dr.Alireza Talebpour, Seyed Hossein Khasteh
Dr. Nasser Mozayan

 زمان ۱۰ دی ماه ۱۴۰۴

  ساعت: ۱۷:۰۰
 

مکان: اتاق دفاع دکتری
 

چکیده

This research introduces a model that can be used to improve the named entity recognition (NER) of Arabic text with particular reference to hadiths of Islam. To enhance entity recognition, sequence-to-sequence models to give more exact and thorough text analysis. Using the Noor Hadith dataset, which contains ۵۹,۴۳۰ hadiths, a BIO tagging scheme was developed, where the text was cleaned, stop words were removed, and segmentation was performed. A right-skewed distribution was demonstrated using the analysis of text length as shorter texts were more common. By selecting four models, namely AraBERT, BiLSTM, CNN-BiLSTM Hybrid and AraBERT-LSTM Hybrid, words in the dataset were classified into eight entity categories: person, imam, location, narrator, book, tribe, date, and event. Performance evaluation using precision, recall and F۱ score of four models: AraBERT, LSTM, hybrid CNN-BiLSTM and hybrid AraBERT-LSTM model, the latter hybrid model achieved accuracy of ۰.۹۸۱. The hybrid NER model developed in this study also shows great promise for natural language processing in Arabic. We obtain increased precision and consistency in our hybrid model through applying various modeling strategies, which contribute to more successful resolution of difficult tasks. The performance of the hybrid model on the Arabic Islamic texts of the Hadith has proven to be successful, which opens up some exciting future applications in the natural language processing and similar research. The research
forms a valuable source of other attempts to improve such models and make them applicable to other Arabic text collections.

: Abstract

This research introduces a model that can be used to improve the named entity recognition (NER) of Arabic text with particular reference to hadiths of Islam. To enhance entity recognition, sequence-to-sequence models to give more exact and thorough text analysis. Using the Noor Hadith dataset, which contains ۵۹,۴۳۰ hadiths, a BIO tagging scheme was developed, where the text was cleaned, stop words were removed, and segmentation was performed. A right-skewed distribution was demonstrated using the analysis of text length as shorter texts were more common.
By selecting four models, namely AraBERT, BiLSTM, CNN-BiLSTM Hybrid and AraBERT-LSTM Hybrid, words in the dataset were classified into eight entity categories: person, imam, location, narrator, book, tribe, date, and event. Performance evaluation using precision, recall and F۱ score of four models: AraBERT, LSTM, hybrid CNN-BiLSTM and hybrid AraBERT-LSTM model, the latter hybrid model achieved accuracy of ۰.۹۸۱. The hybrid NER model developed in this study also shows great promise for natural language processing in Arabic.
We obtain increased precision and consistency in our hybrid model through applying various modeling strategies, which contribute to more successful resolution of difficult tasks. The performance of the hybrid model on the Arabic Islamic texts of the Hadith has proven to be successful, which opens up some exciting future applications in the natural language processing and similar research. The research forms a valuable source of other attempts to improve such models and make them applicable to other Arabic text collections.