Harnessing the CRF Complexity with Domain-Specific Constraints. The Case of Morphosyntactic Tagging of a Highly Inflected Language

Proceedings of COLING 2012: Technical Papers, pages 2789–2804, COLING 2012, Mumbai, December 2012. Harnessing the CRF complexity with domain-specific constraints. The case of morphosyntactic tagging of a highly inflected language. Jakub Waszczuk Institute of Computer Science, Polish Academy of Sciences ❏❛❦✉❜✳❲❛s③❝③✉❦❅✐♣✐♣❛♥✳✇❛✇✳♣❧ ABSTRACT We describe a domain-specific method of adapting conditional random fields (CRFs) to mor- phosyntactic tagging of highly-inflectional languages. The solution involves extending CRFs with additional, position-wise restrictions on the output domain, which are used to impose con- sistency between the modeled label sequences and morphosyntactic analysis results both at the level of decoding and, more importantly, in parameters estimation process. We decompose the problem of morphosyntactic disambiguation into two consecutive stages of the context-sensitive morphosyntactic guessing and the disambiguation proper. The division helps in designing well-adjusted, CRF-based methods for both tasks, which in combination constitute Concraft, a highly accurate tagging system for the Polish language available under the 2-clause BSD license. Evaluation on the National Corpus of Polish shows that our solution significantly outperforms other state-of-the-art taggers for Polish – Pantera, WMBT and WCRFT – especially in terms of the accuracy measured with respect to unknown words. KEYWORDS: morphosyntactic tagging, conditional random fields, Polish. 2789 1 Introduction Morphosyntactic tagging is one of the basic problems in Natural Language Processing (NLP), which can be described as choosing most appropriate morphosyntactic descriptions for each word in the particular sentence. We ignore the problem of lemma-ambiguity in our work, since it doesn’t occur often within the context of highly-inflectional languages like Polish or Czech (Smith et al., 2005), and concentrate on the problem of choosing correct values of grammatical attributes. Numerous methods have been proposed as a solution to the tagging problem, ranging from rule-based with hand-crafted rules, through transformation-based with rules automatically extracted from a dataset, to stochastic methods like hidden Markov models (Jurafsky and Martin, 2008) or linear-chain conditional random fields (Lafferty et al., 2001; Sutton and Mccallum, 2006). We focus on stochastic methods in this work. Best solutions for the English language yield tagging accuracy exceeding 97%, but methods designed for one language do not necessarily solve problems inherent to another one. In case of highly-inflectional languages the first (and rather technical) obstacle is the complexity of the morphosyntactic tagset. Polish tags carry information about a number of grammatical attributes – part of speech (POS), case, number, gender etc. – with 13 grammatical categories in total. The Polish tagset used within the 1-million-word National Corpus of Polish (NCP) (Przepiórkowski et al., 2010) is a positional one: first position of the tag represents the grammatical class and subsequent positions contain category values. In the NCP tagset the first position determines the set of grammatical categories (some obligatory, some optional) used in conjunction with the class. For example, the subst:sg:nom:m1 tag describes a singular (sg) human masculine (m1) noun (subst – substantive) in the nominative case (nom). There are more than 1000 morphosyntactic tags which can be directly extracted from the NCP corpus. It makes the application of methods, which …

Content truncated to 500 words