Assigning Fine-grained PoS Tags based on High-precision Coarse-grained Tagging

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 328–336, Osaka, Japan, December 11-17 2016. Assigning Fine-grained PoS Tags based on High-precision Coarse-grained Tagging Tobias Horsmann and Torsten Zesch Language Technology Lab Department of Computer Science and Applied Cognitive Science University of Duisburg-Essen, Germany {tobias.horsmann,torsten.zesch}@uni-due.de Abstract We propose a new approach to PoS tagging where in a first step, we assign a coarse-grained tag corresponding to the main syntactic category. Based on this high-precision decision, in the sec- ond step we utilize specially trained fine-grained models with heavily reduced decision complex- ity. By analyzing the system under oracle conditions, we show that there is a quite large potential for significantly outperforming a competitive baseline. When we take error-propagation from the coarse-grained tagging into account, our approach is on par with the state of the art. Our approach also allows tailoring the tagger towards recognizing single word classes which are of interest e.g. for researchers searching for specific phenomena in large corpora. In a case study, we significantly outperform a standard model that also makes use of the same optimizations. 1 Introduction When a part-of-speech (PoS) tagger assigns word class labels to tokens, it has to select from a set of possible labels whose size usually ranges from fifty to several hundred labels depending on the language. Especially for new domains or under-resourced languages, there is usually not enough training data to reliably learn all the subtle differences between a large set of labels. We thus propose to split the PoS tagging task into two steps. A first high precision step, where we only assign a coarse-grained tag which can be reliably learned also with rather limited training data, and which benefits from additional unlabeled data in the form of clusters or embeddings. And a second step, where we apply specialized tagging models that only have to choose from a much smaller set of possible labels. Figure 1 gives an overview of our approach using, in the first step, a coarse-grained tagset (similar to the universal tagset (Petrov et al., 2012)) and in the second step the fine-grained PTB tagset (Marcus et al., 1993). In the figure, we can see how knowing the coarse-grained tag informs the second step. For example, if we already know that beautiful is an adjective, we only have to choose between three possible tags (JJ, JJR, or JJS) instead of 45 tags for the full PTB tagset. Our approach requires that the coarse tagging in the first step is more accurate than using fine-grained tagging itself, as we will loose some accuracy through error propagation between the steps. We will thus first analyze how well coarse tagging can actually be done, and also focus on whether coarse-grained models transfer better between different kinds of texts, as e.g. Ritter et al. (2011) shows that a fine- grained tagger trained on newswire data doesn’t work well on social media. Creating robust and accurate coarse-grained taggers is a worthwhile task on its own, as …

Content truncated to 500 words