Urdu Dependency Parser: A Data-Driven approach

In this paper, we present what we believe to be the first data-driven dependency parser for Urdu. The parser was trained and tuned using MaltParser system, a system for data-driven dependency parsing. The Urdu dependency treebank (UDT) is used for training and testing of the Urdu dependency parser, is also presented first time. The UDT contains corpus of 2853 sentences which are annotated at multiple levels such as part-of-speech (POS) level, chunk (phrase level) and dependency relations level. The UDT also contains information about the token counter, head of current token. The annotation is done manually to build UDT. Urdu Dependency Parsing system is evaluated by conducting a series of experiments. All experiments are performed using Maltparser default algorithm with different feature models. Initial, a base line simple feature model consisting word position, word, head and dependency relation is used for Urdu dependency parsing. Then feature model is enhanced by adding part-of-speech (POS) and chunk (Phrase level) information. The results of all parsing experiments are reported. The overall best labeled accuracy (LA) achieved 74.48% and 90.14% of unlabeled attachment score (UAS) is achieved. The error analysis is performed by comparing output data with treebank test data which manual parsed to analyze and classify the different types of errors produced by the parser. This is very useful to identify the future directions for future expansion of the treebank and for improving the parsing accuracy.