Dependency Parsing for Urdu: Resources, Conversions and Learning

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5202–5207 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5202 Dependency Parsing for Urdu: Resources, Conversions and Learning Toqeer Ehsan1, Miriam Butt2 1Department of Computer Science, University of Gujrat, Pakistan 2Department of Linguistics, University of Konstanz, Germany 1toqeer.ehsan@uog.edu.pk, 2miriam.butt@uni-konstanz.de Abstract This paper adds to the available resources for the under-resourced language Urdu by converting different types of existing treebanks for Urdu into a common format that is based on Universal Dependencies. We present comparative results for training two dependency parsers, the MaltParser and a transition-based BiLSTM parser on this new resource. The BiLSTM parser incorporates word embeddings which improve the parsing results significantly. The BiLSTM parser outperforms the MaltParser with a UAS of 89.6 and an LAS of 84.2 with respect to our standardized treebank resource. Keywords: Urdu, Dependency Treebank, Parsing 1. Introduction In this paper, we tackle the lack of resources for Urdu statistical parsing by converting different types of exist- ing treebanks for Urdu into a common format that is based on the Universal Dependency (UD) 2.0 label set (Nivre et al., 2017). We use this combined new resource to train and comparatively evaluate two state-of-the-art dependency parsers, namely the MaltParser (Nivre et al., 2007) and a transition-based BiLSTM parser (Kiperwasser and Gold- berg, 2016). In addition, we experiment with the incorpo- ration of word embeddings and find that this significantly improves the parsing results. A Phrase structure (PS) treebank represents the con- stituency of a clause, the linear order and the hierarchi- cal organization of the constituents. Information about a predicate’s arguments are encoded only indirectly and their immediate accessibility depends on the precise type of PS treebank. In contrast, dependency structures (DS) abstract away from linear order and concentrate on encoding func- tional dependencies between the items of a clause. This mainly encompasses, but is not restricted to, grammati- cal or semantic relations between a predicate and its argu- ments. Grammatical relations are important in applications of nat- ural language understanding (NLU) as they provide infor- mation on event participants. One can include functional labels representing grammatical relations in a PS parser. However, it has been shown that training a PS parser by in- cluding functional labels produces lower constituency pars- ing accuracy. The Stanford parser (Klein and Manning, 2003) together with functional labels produces a functional accuracy of 68.3% on a PS Urdu treebank. On the other hand, dependency parsers predict dependency labels with higher accuracies to represent grammatical relations. Both types of treebanks of varying size already exist for Urdu and we conclude that a promising way forward for train- ing high-quality dependency parsers is to convert existing PS treebanks into equivalent DS treebanks, rather than to enhance an existing PS treebank with functional labels. In this paper, we thus pool existing treebank resources for Urdu and convert an existing PS treebank to a DS format that uses the UD label set. In order to achieve …

Content truncated to 500 words