Dependency Parsing and Domain Adaptation with Data-Driven LR Models and Parser Ensembles

Abstract

We present a data-driven variant of the LR algorithm for dependency parsing, and extend it with a best-first search for probabilistic generalized LR dependency parsing. Parser actions are determined by a classifier, based on features that represent the current state of the parser. We apply this parsing framework to both tracks of the CoNLL 2007 shared task, in each case taking advantage of multiple models trained with different learners. In the multilingual track, we train three LR models for each of the ten languages, and combine the analyses obtained with each individual model with a maximum spanning tree voting scheme. In the domain adaptation track, we use two models to parse unlabeled data in the target domain to supplement the labeled out-ofdomain training set, in a scheme similar to one iteration of co-training.

Experiments in Newswire-to-Law Adaptation of Graph-Based Dependency Parsers

“…Somewhat surprisingly very few of the methods that have been previously proposed in the literature seem to be efficient on the evaluation campaign data set, including [9,5,8,10]. Some of our experiments led to significant or nearsignificant improvements on development data, but the same set-ups led to poor results on test data.…”

“…Strategies to automatically correct sample bias in natural language processing include feature-based approaches [1,4], instance weighting [3,8,10] and using semi-supervised learning algorithms [9,2]. Most attempts to use feature-based approaches in parsing have failed, and in our experiments we therefore focused on instance weighting and semi-supervised learning algorithms.…”

“…Co-training This approach is inspired by [9]. Training our base parser on all sentences the MSTParser and the MaltParser agree on (exact matches) as well as the original source data never led to improvements over our base parser, but using MSTParser and Mate-Tools led to small improvements.…”

“…We evaluate two very different methods for domain adaptation of graph-based dependency parsers on the EVALITA 2011 Domain Adaptation data, namely instance-weighting [10] and self-training [9,6]. Since the source and target domains (newswire and law, respectively) were very similar, instance-weighting was unlikely to be efficient, but some of the semi-supervised approaches led to significant improvements on development data.…”

Dependency Chart Parsing Algorithm Based on Ternary-Span Combination

“…[3] 93.04% -Chen [19] 92.21% -Sagae [20] -88.94% Nakagawa [15] -88.88% high-order [7] 93.14% b 86.20%…”