FlexTag: A Highly Flexible PoS Tagging Framework

FlexTag: A Highly Flexible PoS Tagging Framework Torsten Zesch, Tobias Horsmann Language Technology Lab Department of Computer Science and Applied Cognitive Science University of Duisburg-Essen, Germany {torsten.zesch,tobias.horsmann}@uni-due.de Abstract We present FlexTag, a highly flexible PoS tagging framework. In contrast to monolithic implementations that can only be retrained but not adapted otherwise, FlexTag enables users to modify the feature space and the classification algorithm. We categorize existing PoS tagger implementations into one of three categories with regards to model-training capabilities and the level of access those implementations give a researcher to intrinsic details. To this categorization we add a new fourth layer characterizing our FlexTag which fully exposes the feature space and the machine learning algorithm while sustaining a high usability. With FlexTag, rapid prototyping of tagger models using different feature spaces can be easily implemented, taking a huge technical burden from the NLP researcher who is experimenting with new ideas or resources. FlexTag makes it easy to quickly develop custom-made taggers exactly fitting a research problem. We demonstrate the capabilities of FlexTag by first training a PoS tagger model with state-of-the-art performance on the common Wall-Street-Journal split using a default feature set. Furthermore, we train a social media model fitted to the Twitter domain that extends the default feature set by adding domain-specific features and incorporating knowledge from unsupervised data resources. Keywords: PoS tagging, domain adaptation, feature engineering 1. Motivation Part-of-speech tagging is an important preprocessing step in natural language processing. Consequently, there are many implementations available such as the Stanford tag- ger (Toutanova et al., 2003), the TreeTagger (Schmid, 1994), or the ClearNLP tagger (Choi and Palmer, 2012). In order to adapt the behaviour of the tagger to the needs of the user, most taggers provide configurable models which are targeted towards different languages or domains. A user might e.g. switch between a standard English model and a German social media model depending on the task at hand. As there rarely is a perfect match between the available tag- ging models and the documents to be processed, most tag- gers allow users to train their own models. FlexTag goes beyond the usual re-training possibilities of established tag- gers by giving the user flexible control over the feature extraction and classification step. Users can choose from different machine learning algorithms or plug-in their own learning framework. FlexTag already comes with a wide range of implemented feature extraction modules that can be enabled when needed. However, users can also write new feature extraction modules. For example, unsuper- vised information from Brown clusters (Ritter et al., 2011) or LDA (Rehbein, 2013) is known to boost performance quite a bit, but the corresponding features are usually not supported in standard taggers. FlexTag is embedded in an evaluation framework (Hors- mann et al., 2015) that allows for quick comparison of the newly created taggers with previous versions and other im- plementations. 2. FlexTag Architecture In order to better understand in which way FlexTag differs from existing PoS taggers, we distinguish four commonly used architectures …

Content truncated to 500 words