In this project, we focused on the analysis of Hebrew sentences, with the overall purpose of researching and developing a unique model capable of identifying the subject in sentences. Identifying a subject is quite complex, on the one hand, and of great importance, on the other. Among the many important benefits that can be gained from such a system is the ability to optimize systems of categorization, summarization and translation. To tackle this task, we used deep learning technologies and research methods. After
researching, we built models from three types of complex neuronal networks -- Bi- LSTM, LSTM and GRU -- with the model architecture being Seq2Seq combined with Attention Mechanism. For Word Embedding we used Word2Vec.
The results of the study achieved a high percentage of success. In terms of model fit and Loss function, Bi-LSTM with Attention showed 97.65 percent matching results, the LSTM model with Attention presented 0.82 and the GRU with Attention 0.98 percent. These results were significantly superior to classical models that exist today in the field.
Corpus: https://github.com/NLPH/SVLM-Hebrew-Wikipedia-Corpus/blob/master/SVLM_Hebrew_Wikipedia_Corpus.txt
As it was generated from Hebrew Wikipedia sources, which are licensed under the CC-BY-SA 3.0_ license, this corpus is thus also necessarilly licensed under the same license.
CC-BY-SA 3.0: https://creativecommons.org/licenses/by-sa/3.0/