Neural Dependency Parsıng For Turkısh

Abstract

Dependency Parsing is the task of finding the grammatical structure of a sentence by identify- ing syntactic and semantic relationships between words. The current accuracy of the depen- dency parsers is still not satisfying due to the long term dependencies and out-of-vocabulary (OOV) problem. Those problems also apply to Turkish because of the high percentage of OOV words due to its agglutinative morphological structure compared to other languages.

The recent work shows that Recurrent Neural Networks (RNNs) are not efficient for long sequences. The deep neural architecture that we propose in this thesis follows an encoder- decoder structure with an encoder based on a Transformer Network and a decoder based on a Stack Pointer Network. The character-level word embeddings are also integrated in the model to cope with the OOV problem. The results for both Turkish and English show that the proposed model performs better for long sentences and can identify long term dependencies more efficiently compared to other neural models.