Coupling an Annotated Corpus and a Morphosyntactic Lexicon for State-of-the-Art POS Tagging with Less Human Effort

Coupling an Annotated Corpus and a Morphosyntactic Lexicon for State-of-the-Art POS Tagging with Less Human Effort Pascal Denis and Benoît Sagot Equipe-project ALPAGE INRIA and Université Paris 7 30, rue Château des Rentiers 75013 Paris, France {pascal.denis,benoit.sagot}@inria.fr Abstract. This paper investigates how to best couple hand-annotated data with information extracted from an external lexical resource to improve POS tagging performance. Focusing on French tagging, we introduce a maximum entropy conditional sequence tagging system that is enriched with information extracted from a morphological resource. This system gives a 97.7% accuracy on the French Treebank, an error reduction of 23% (28% on unknown words) over the same tagger without lexical information. We also conduct experiments on datasets and lexicons of varying sizes in order to assess the best trade-off between annotating data vs. developing a lexicon. We find that the use of a lexicon improves the quality of the tagger at any stage of development of either resource, and that for fixed performance levels the availability of the full lexicon consistently reduces the need for supervised data by at least one half. Keywords: Part-of-speech tagging, maximum entropy models, morphosyntactic lexicon, French, language resource development 1 Introduction Over recent years, numerous systems for automatic part-of-speech (POS) tagging have been pro- posed for a large variety of languages. Among the best performing systems are those based on supervised machine learning techniques (see (Manning and Schütze, 1999) for an overview). For some languages like English and other European languages, these systems have reached perfor- mance that comes close to human levels. Interestingly, the majority of these systems have been built without resorting to any external lexical information sources; they instead rely on a dictio- nary that is based on the training corpus (see however (Hajiˇc, 2000)). This raises the question of whether we can still improve tagging performance by exploiting this type of resource. Arguably, a potential advantage of using an external dictionary is in a better handling of unknown words (i.e., words that are not present in the training corpus, but that may be present in the external dictio- nary). A subsequent question is how to best integrate the information from a lexical resource into a probabilistic POS tagger. In this paper, we consider two distinct scenarios: (i) using the external dictionary as constraints that restrict the set of possible tags that the tagger can choose from, and (ii) incorporating the dictionary tags as features in a probabilistic POS tagging model. Another in- teresting question is that of the relative impact of training corpora and of lexicons of various sizes. This issue is crucial to the development of POS taggers for resource-scarce languages for which it is important to determine the best trade-off between annotating data and constructing dictionaries. This paper addresses these questions through various tagging experiments carried out on French, based on our new tagging system called MElt (Maximum-Entropy Lexicon-enriched Tag- ger). An obvious motivation for working on this language is the availability of a training corpus Copyright 2009 by Pascal Denis …

Content truncated to 500 words