Training Parsers on Incompatible Treebanks

Proceedings of NAACL-HLT 2013, pages 127–137, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Training Parsers on Incompatible Treebanks Richard Johansson Spr˚akbanken, Department of Swedish, University of Gothenburg Box 200, SE-40530 Gothenburg, Sweden richard.johansson@gu.se Abstract We consider the problem of training a sta- tistical parser in the situation when there are multiple treebanks available, and these tree- banks are annotated according to different lin- guistic conventions. To address this problem, we present two simple adaptation methods: the first method is based on the idea of using a shared feature representation when parsing multiple treebanks, and the second method on guided parsing where the output of one parser provides features for a second one. To evaluate and analyze the adaptation meth- ods, we train parsers on treebank pairs in four languages: German, Swedish, Italian, and En- glish. We see significant improvements for all eight treebanks when training on the full training sets. However, the clearest benefits are seen when we consider smaller training sets. Our experiments were carried out with unlabeled dependency parsers, but the meth- ods can easily be generalized to other feature- based parsers. 1 Introduction When developing a data-driven syntactic parser, we need to fit the parameters of its statistical model on a collection of syntactically annotated sentences – a treebank. Generally speaking, a larger collection of examples in the training treebank will give a higher quality of the resulting parser, but the cost in time and effort of annotating training sentences is fairly high. Most existing treebanks are in the range of a few thousand sentences. However, there is an abundance of theoretical models of syntax and there is no consensus on how treebanks should be annotated. For some languages, there exist multiple treebanks annotated according to different syntactic theories. Apart from German, Swedish, and Italian, which will be considered in this paper, there are important examples among the world’s major languages, such as Arabic and Chi- nese. To exemplify how syntactic annotation conven- tions may differ in even such a simple case as un- labeled dependency annotation, consider the Italian sentence fragment la sospensione o l’interruzione (’the suspension or the interruption’) in Figure 1. As we will see in detail in §3.1.3, there are two Ital- ian treebanks: the ISST and TUT. If annotating as in the ISST treebank (drawn above the sentence) determiners (la, l’) are annotated as dependents of the following nouns (sospensione, interruzione); in TUT (drawn below the sentence), we have the re- verse situation. There are also differences in how coordinate structures are represented: in ISST, the two conjuncts are directly conjoined and the con- junction attached to the first of them, while in TUT the conjunction acts as a link between the conjuncts. o lasospensione interruzione l’ Figure 1: Differences in dependency annotation styles. Given the high cost of treebank annotation and the importance of a proper amount of data for parser de- velopment, this situation is frustrating. How could we then make use of multiple treebanks when train- …

Content truncated to 500 words