Data-Driven Dependency Parsing of New Languages Using Incomplete and Noisy Training Data

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 12–20, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Data-Driven Dependency Parsing of New Languages Using Incomplete and Noisy Training Data Kathrin Spreyer and Jonas Kuhn Department of Linguistics University of Potsdam, Germany {spreyer,kuhn}@ling.uni-potsdam.de Abstract We present a simple but very effective ap- proach to identifying high-quality data in noisy data sets for structured problems like parsing, by greedily exploiting partial struc- tures. We analyze our approach in an anno- tation projection framework for dependency trees, and show how dependency parsers from two different paradigms (graph-based and transition-based) can be trained on the result- ing tree fragments. We train parsers for Dutch to evaluate our method and to investigate to which degree graph-based and transition- based parsers can benefit from incomplete training data. We find that partial correspon- dence projection gives rise to parsers that out- perform parsers trained on aggressively fil- tered data sets, and achieve unlabeled attach- ment scores that are only 5% behind the aver- age UAS for Dutch in the CoNLL-X Shared Task on supervised parsing (Buchholz and Marsi, 2006). 1 Introduction Many weakly supervised approaches to NLP rely on heuristics or filtering techniques to deal with noise in unlabeled or automatically labeled training data, e.g., in the exploitation of parallel corpora for cross- lingual projection of morphological, syntactic or se- mantic information. While heuristic approaches can implement (linguistic) knowledge that helps to de- tect noisy data (e.g., Hwa et al. (2005)), they are typ- ically task- and language-specific and thus introduce a component of indirect supervision. Non-heuristic filtering techniques, on the other hand, employ re- liability measures (often unrelated to the task) to predict high-precision data points (e.g., Yarowsky et al. (2001)). In order to reach a sufficient level of precision, filtering typically has to be aggressive, especially for highly structured tasks like parsing. Such aggressive filtering techniques incur massive data loss and enforce trade-offs between the quality and the amount of usable data. Ideally, a general filtering strategy for weakly su- pervised training of structured analysis tools should eliminate noisy subparts in the automatic annota- tion without discarding its high-precision aspects; thereby data loss would be kept to a minimum. In this paper, we propose an extremely simple ap- proach to noise reduction which greedily exploits partial correspondences in a parallel corpus, i.e., correspondences potentially covering only substruc- tures of translated sentences. We implemented this method in an annotation projection framework to create training data for two dependency parsers rep- resenting different parsing paradigms: The MST- Parser (McDonald et al., 2005) as an instance of graph-based dependency parsing, and the Malt- Parser (Nivre et al., 2006) to represent transition- based dependency parsing. In an empirical evalu- ation, we investigate how they react differently to incomplete and noisy training data. Despite its simplicity, the partial correspondence approach proves very effective and leads to parsers that achieve unlabeled attachment scores that are only 5% behind the average UAS for Dutch in the CoNLL-X …

Content truncated to 500 words