Feature engineering using shallow parsing in argument classification of Persian verbs

Identifying the verb´s dependents and determining the semantic role for them is a natural pre-processing step in applications such as machine translation (MT) and question answering (QA). In this paper, we present a feature set for assigning argument instances into thematic role classes such as “Agent” and “Patient”. This feature set contains mainly language specific features for syntactic segments (chunks) of Persian sentences which can be categorized into three feature types including verb properties, chunk content and relation between the argument and verb of a sentence. We train an instance-based classifier on our manually annotated dataset to select the appropriate semantic role of each chunk. The classifier discriminates the best semantic role without considering the interaction between chunks in a sentence. The results show that our feature set discriminates the thematic roles of arguments in a considerable accuracy about 81.9% which enhances the baseline accuracy about 18.8%. Our dataset is free release and available for the researchers.