M.S Thesis: Transition Based Dependency Parsing with Deep Learning, Koç University, Department of Computer Engineering. September 2018. (PDF, Presentation).
Publications: CoNLL18 and CoNLL17
Code: CoNLL18 and CoNLL17
Thesis Abstract:
I introduce word and context embeddings derived from a language model representing left/right context of a word instance and demonstrate that context embeddings significantly improve the accuracy of transition based parser. Our multi-layer perceptron (MLP) parser making use of these embeddings was ranked 7th out of 33 participants (ranked 1st among transition based parsers) in CoNLL 2017 UD Shared Task. However MLP parser relies on additional hand-crafted features which are used to summarize sequential information. I exploit recurrent neural networks to remove these features by implementing tree-stack LSTM, and develop new set of continuous embeddings called morphological feature embeddings. According to official comparison results in CoNLL 2018 UD Shared Task, our tree-stack LSTM outperforms MLP in transition based dependency parsing.
Full post...