LTAG-spinal treebank and parser for Hindi

Statistical parsers need huge annotated treebanks to learn from and building treebanks is an expensive proposition. To create parsers for different grammar formalisms in a language, building separate treebanks for each of those isn’t a feasible task. Treebanks available in one formalism can be converted into an other either automatically or with minimal human effort by exploiting the similarities and differences between the two. In this work, we present an approach to extract an LTAGspinal treebank from Hyderabad Dependency Treebank for Hindi. LTAG-spinal is a variant of Lexicalized Tree Adjoining Grammar (LTAG) with desirable linguistic, computational and statistical properties. A bidirectional LTAG dependency parser is trained on the extracted treebank and an LTAG dependency accuracy of 80.86% is reported.