Syntactic Parsing and Compound Recognition via Dual Decomposition: Application to French

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1875–1885, Dublin, Ireland, August 23-29 2014. Syntactic Parsing and Compound Recognition via Dual Decomposition: Application to French Joseph Le Roux1 and Matthieu Constant2 and Antoine Rozenknop1 (1) LIPN, Université Paris 13 – Sorbonne Paris Cité, CNRS (2) LIGM, Université Paris Est, CNRS leroux@univ-paris13.fr, mconstan@univ-mlv.fr, antoine.rozenknop@lipn.univ-paris13.fr Abstract In this paper we show how the task of syntactic parsing of non-segmented texts, including com- pound recognition, can be represented as constraints between phrase-structure parsers and CRF sequence labellers. In order to build a joint system we use dual decomposition, a way to com- bine several elementary systems which has proven successful in various NLP tasks. We evaluate this proposition on the French SPMRL corpus. This method compares favorably with pipeline architectures and improves state-of-the-art results. 1 Introduction Dual decomposition (DD), which can be used as a method to combine several elementary systems, has already been successfully applied to many NLP tasks, in particular syntactic parsing, see (Rush et al., 2010; Koo et al., 2010) inter alia. Intuitively, the principle can be described quite simply: at decoding time, the combined systems seek for a consensus on common subtasks, in general the prediction of some parts of the overall structure, via an iterative process imposing penalties where the systems disagree. If the systems converge to a solution, it is formally guaranteed to be optimal. Besides, this approach is quite flexible and easy to implement. One can add or remove elementary systems without rebuilding the architecture from the ground up. Moreover, the statistical models for the subsystems can often be estimated independently at training time. In this paper we show how syntactic parsing of unsegmented texts, integrating compound recognition, can be represented by constraints between phrase-structure parsers and sequence labellers, either for compound recognition or part-of-speech (POS) tagging, and solved using DD. We compare this approach experimentally with pipeline architectures: our system demonstrates state-of-the-art performance. While this paper focuses on French, the approach is generic and can be applied to any treebank with compound information, and more generally to tasks combining segmentation and parsing. This paper is structured as follows. First, we describe the data we use to build our elementary systems. Second, we present related work in compound recognition, in particular for French, and the type of information one is able to incorporate in tag sets. Third, we show how CRF-based sequence labellers with these different tag sets can be combined using DD to obtain an efficient decoding algorithm. Fourth, we extend our method to add phrase-structure parsers in the combination. Finally, we empirically evaluate these systems and compare them with pipeline architectures. 2 Data We use the phrase-structure treebank released for the SPMRL 2013 shared task (Seddah et al., 2013). This corresponds to a new version of the French Treebank (Abeillé et al., 2003). One of the key dif- ferences between French data and other treebanks of the shared task is the annotation of compounds. Compounds are sequences of …

Content truncated to 500 words