Semi-Supervised Training for the Averaged Perceptron POS Tagger

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 763–771, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Semi-supervised Training for the Averaged Perceptron POS Tagger Drahom´ıra “johanka” Spoustov´a Jan Hajiˇc Jan Raab Institute of Formal and Applied LinguisticsMiroslav Spousta Faculty of Mathematics and Physics, Charles University Prague, Czech Republic {johanka,hajic,raab,spousta}@ ufal.mff.cuni.cz Abstract This paper describes POS tagging exper- iments with semi-supervised training as an extension to the (supervised) averaged perceptron algorithm, first introduced for this task by (Collins, 2002). Experiments with an iterative training on standard-sized supervised (manually annotated) dataset (106 tokens) combined with a relatively modest (in the order of 108 tokens) un- supervised (plain) data in a bagging-like fashion showed significant improvement of the POS classification task on typo- logically different languages, yielding bet- ter than state-of-the-art results for English and Czech (4.12 % and 4.86 % relative er- ror reduction, respectively; absolute accu- racies being 97.44 % and 95.89 %). 1 Introduction Since 2002, we have seen a renewed interest in improving POS tagging results for English, and an inflow of results (initial or improved) for many other languages. For English, after a relatively big jump achieved by (Collins, 2002), we have seen two significant improvements: (Toutanova et al., 2003) and (Shen et al., 2007) pushed the results by a significant amount each time.1 1In our final comparison, we have also included the re- sults of (Gim´enez and M`arquez, 2004), because it has sur- passed (Collins, 2002) as well and we have used this tag- ger in the data preparation phase. See more details below. Most recently, (Suzuki and Isozaki, 2008) published their Semi-supervised sequential labelling method, whose results on POS tagging seem to be optically better than (Shen et al., 2007), but no significance tests were given and the tool is not available for download, i.e. for repeating the results and sig- nificance testing. Thus, we compare our results only to the tools listed above. Even though an improvement in POS tagging might be a questionable enterprise (given that its effects on other tasks, such as parsing or other NLP problems are less than clear—at least for En- glish), it is still an interesting problem. Moreover, the “ideal”2 situation of having a single algorithm (and its implementation) for many (if not all) lan- guages has not been reached yet. We have cho- sen Collins’ perceptron algorithm because of its simplicity, short training times, and an apparent room for improvement with (substantially) grow- ing data sizes (see Figure 1). However, it is clear that there is usually little chance to get (substan- tially) more manually annotated data. Thus, we have been examining the effect of adding a large monolingual corpus to Collins’ perceptron, appro- priately extended, for two typologically different languages: English and Czech. It is clear however that the features (feature templates) that the tag- gers use are still language-dependent. One of the goals is also to have a fast im- plementation for tagging large amounts …

Content truncated to 500 words