چكيده به لاتين
Semantic Role Labeling is one of the fundamental issues in natural language processing and a key step towards understanding natural language, and is responsible for automatically answering questions such as "who, did what, to whom, when, where, why, and how". In recent years researchers have focused on designing Cross-Lingual methods for Semantic Role Labeling. In Cross-Lingual Semantic Role Labeling we want to use the resources of a language like English that has many high-quality samples to improve the results in low-resource language. Various approaches have been proposed for cross-lingual semantic role labeling, including annotation projection, translation-based approaches, and model transfer. In this research, a deep learning algorithm based on model transfer is proposed, in which the structure of the model allows it to be trained on multiple languages simultaneously and uses data from different languages to optimize the network. We use the English part of the CoNLL2009 multilingual dataset and the corpus of Persian Proposition Bank. In the proposed model, we used only 10% of the training dataset instead of using the entire dataset. Following previous research, we followed the usual division of the training, validation, and testing datasets in English, and used 10% test data, 10% validation data, and 80% training data settings in Persian, resulting in an F1 score of 71.76 in monolingual mode and 74.11 in cross-lingual mode with 10% English data. Additionally, if we use all available English data, the F1 score reaches 75.94. Therefore, this model improved Persian results by 4.18% compared to the monolingual mode, by utilizing English data. By conducting similar experiments using Nicksirat's model, which is the best semantic labeling model in Persian, we achieved an F1 score of 69.71. Therefore, our proposed model improved by 2.05% in monolingual mode and 6.23% in cross-lingual mode compared to this model. However, since our model performs four stages of semantic labeling, unlike Nicksirat's research that assumes the first two stage are given and uses the gold data, the actual difference between the two models is significantly higher than 6.23%.
