Persian Semantic Role Labeling Based on Dependency Tree

Semantic role labeling is the task of attaching semantic tags to the words according to the event represented by the sentence. Persian semantic role labeling is a challenging task and most methods proposed so far depend on a huge number of manually extracted features and are applied on feature engineering to attain high performance. On the other hand, considering the Free-Word-Order and Subject-Object-Verb-Order characteristics of Persian, the arguments of the verbal predicate are often distant and create long-range dependencies. The long-range dependencies can hardly be modeled by these methods. Our goal is to achieve a better performance only with minimal feature engineering and also to capture long-range dependencies in a sentence. To these ends, in this paper a deep model for semantic role labeling is developed with the help of dependency tree for Persian. In our proposed method, for each verbal predicate, the potential arguments are identified by dependency relations, and then the dependency path for each pair of predicate and its candidate argument is embedded using the information in the dependency trees. In the next step, we employed a bi-directional recurrent neural network with long short-term memory units to transform word features into semantic role scores. Experiments have been done on the First Semantic Role Corpus in Persian Language and the corpus provided by the authors. The achieved Macro-average F1-measure is 80.01 for the first corpus and 82.48 for the second one