Ensemble Romanian Dependency Parsing with Neural Networks

Ensemble Romanian Dependency Parsing with Neural Networks Radu Ion, Elena Irimia, Verginica Barbu Mititelu Research Institute for Artificial Intelligence, “Mihai Drăgănescu” 13 “Calea 13 Septembrie”, Bucharest 050711, Romania {radu, elena, vergi}@racai.ro Abstract SSPR (Semantics-driven Syntactic Parser for Romanian) is a neural network ensemble parser developed for Romanian (a Python 3.5 application based on the Microsoft Cognitive Toolkit 2.0 Python API) that combines the parsing decisions of a varying number (in our experiments, 3) of other parsers (MALT, RGB and MATE), using information from additional lexical, morpho-syntactic and semantic features. SSPR outperforms the best individual parser (MATE in our case) with 1.6% LAS points and it is in the same class with the top 5 Romanian performers at the CONLL 2017 dependency parsing shared task. The train and test sets were extracted from a Romanian dependency treebank we developed and validated in the Universal Dependencies format. The treebank, used in the CONLL 2017 Romanian track as well, is open licenced; the parser is available on request. Keywords: parsing, treebank, neural networks. 1. Introduction A reliable, fast and freely available syntactic parser for Romanian was needed (by both linguists and computer scientists working in Natural Language Processing) when we decided to develop SSPR. Even if other attempts at creating Romanian parsers were documented before, none proved to be accessible or performant enough: Călăcean and Nivre (2009) reported a dependency parser based on MaltParser (Nivre et al., 2007) and trained on a small treebank (Hristea and Popescu, 2003), whose impressive results, 88.6% labelled attachment score (LAS) and 92% unlabelled attachment score (UAS), are expectable given the shortness and syntactic simplicity of the sentences in the data set; the Fips constituency rule-based parser was adapted for Romanian (Seretan et al., 2010) and is available for online use (http://www.latl.unige.ch/), but no evaluation is provided for this language; Colhon and Cristea (2016) focus only on noun phrases, using patterns to identify relations inside such structures; the dependency parser of Perez et al. (2016) reported 78.04% LAS when trained on an 8000 sentences treebank. Dumitrescu et al. (2017) RBG-based parser obtained a LAS of 79.44% in the CONLL 2017 Romanian track but new developments of Boroș and Dumitrescu (2018) employ a Bidirectional LSTM network that is still training at the time of the writing and whose foreseen results are much better. In a nutshell, the SSPR ensemble parser we have developed works by training a feed-forward neural network (NN) to recognize “good” vs. “bad” dependency relations from the minimum spanning tree (MST) of the directed graph of all existing (as many as possible) parses for the same sentence. For the training part, each edge of the graph is weighted by the inverse of the size of the set of parsers that found that edge (the “majority voting” combination strategy). The aim of the SSPR ensemble parser is to improve on the baseline “majority voting” combination of dependency parses by using the additional, lexical, morpho-syntactic and semantic features of the words participating in a relation, coupled with information …

Content truncated to 500 words