Arabic text generation : deep learning for poetry synthesis

Hejazi, HD ORCID: https://orcid.org/0000-0001-8726-3660, Khamees, AA ORCID: https://orcid.org/0000-0002-9324-464X, Alshurideh, M ORCID: https://orcid.org/0000-0002-7336-381X and Salloum, SA ORCID: https://orcid.org/0000-0002-6073-3981 2021, Arabic text generation : deep learning for poetry synthesis , in: International Conference on Advanced Machine Learning Technologies and Applications, 20th-22nd March 2021, Cairo, Egypt.

Full text not available from this repository. (Request a copy)

Abstract

Text Generation, especially poetry synthesis, is a promising and challenging AI task. We have used LSTM and word2vec methods to explore this area. We do forward and backward word training with different word sequences lengths. Two Datasets of Arabic poems were used. Preprocessing for unification and cleaning was done too, but the Data size was big and required very high memory and processing, so we used a sub-Datasets for training; this affected our experiments since the model is trained on fewer data. A user-supplied keyword was implemented. We have found the shorter training sequence models were better in generating more meaningful text, and longer models prefer most frequent words, repeat text, and use small words. Best predicted sentences were selected by measuring each of its words conditional probability and multiply them; this avoids local maxima if we used a greedy method that chooses the best next-word only. Moreover, the AraVecword2vec module was not very helpful since it was provided synonyms much more that related words. Many enhancements can be done in the future, such as Arabic prosody constraints, and overcome the hardware issue.

Item Type: Conference or Workshop Item (Paper)
Schools: Schools > School of Computing, Science and Engineering
Journal or Publication Title: Advanced Machine Learning Technologies and Applications : proceedings of AMLTA 2021
Publisher: Springer
Series Name: Advances in Intelligent Systems and Computing
ISBN: 9783030697167 (print); 9783030697174 (ebook)
ISSN: 2194-5357
Related URLs:
Depositing User: USIR Admin
Date Deposited: 22 Jun 2021 12:51
Last Modified: 22 Jun 2021 12:58
URI: http://usir.salford.ac.uk/id/eprint/61006

Actions (login required)

Edit record (repository staff only) Edit record (repository staff only)