Improving Persian Dependency-Based Parser Using Deep Learning

One of the most important problems in computational linguistics is the grammar and, consequently, syntactic structures and structural parsing. The structural parser tries to analyze the relationships between words and to extract the syntactic structure of the sentence. The dependency-based structural parser is proper for free-word-order and morphologically-rich languages such as Persian. The data-driven dependency parser performs the categorization process based on a wide range of features, which, in addition to the problems such as sparsity and curse of dimensionality, it requires the correct selection of the features and proper setting of the parameters. The aim of this study is to obtain high performance with minimal feature engineering for dependency parsing of Persian sentences. In order to achieve this goal, the required features of the Maximum Spanning Tree Parser (MSTParser) are extracted with a Bidirectional Long Short-Term Memory (Bi-LSTM) Network and the edges of the dependency graph is scored by that. Experiments are conducted on the Persian Dependency Treebank (PerDT) and the Uppsala Persian Dependency Treebank (UPDT). The obtained results indicate that the definition of new features improves the performance of the dependency parser for Persian. The achieved unlabeled attachment scores for PerDT and UPDT are 90.53% and 87.02%, respectively.