An Extensive Empirical Evaluation of Character-Based Morphological Tagging for 14 Languages

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 505–513, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics An Extensive Empirical Evaluation of Character-Based Morphological Tagging for 14 Languages Georg Heigold DFKI & Saarland University Saarbr¨ucken, Germany georg.heigold@dfki.de G¨unter Neumann DFKI Saarbr¨ucken, Germany neumann@dfki.de Josef van Genabith DFKI & Saarland University Saarbr¨ucken, Germany josef.van genabith@dfki.de Abstract This paper investigates neural character- based morphological tagging for lan- guages with complex morphology and large tag sets. Character-based approaches are attractive as they can handle rarely- and unseen words gracefully. We eval- uate on 14 languages and observe con- sistent gains over a state-of-the-art mor- phological tagger across all languages ex- cept for English and French, where we match the state-of-the-art. We compare two architectures for computing character- based word vectors using recurrent (RNN) and convolutional (CNN) nets. We show that the CNN based approach performs slightly worse and less consistently than the RNN based approach. Small but sys- tematic gains are observed when combin- ing the two architectures by ensembling. 1 Introduction Character-based approaches have been studied for many applications in natural language pro- cessing, including part-of-speech (POS) tagging (dos Santos and Zadrozny, 2014; Ling et al., 2015; Gillick et al., 2016; Plank et al., 2016; Ma and Hovy, 2016), morphological tagging (Labeau et al., 2015), parsing (Ballesteros et al., 2015), named entity recognition (Gillick et al., 2016), language modeling (Ling et al., 2015; Kim et al., 2016), and neural machine translation (Costa- juss`a and Fonollosa, 2016). Character-based rep- resentations have the advantage of gracefully han- dling rare or unseen words and tend to produce more compact models as the number of atomic units, i.e., characters, is smaller compared to the number of words in word-level approaches. The issue of rare or unseen words is particularly pro- nounced when working on morphologically-rich languages, small amounts of training data or noisy user input. Morphological tagging is the task of assigning a morphological analysis to a token in context. The morphological analysis for a word consists of a sequence of feature:value pairs describing, for example, case, gender, person and tense. A par- ticular concatenation of such feature:value pairs is referred to as a single tag (Oflazer and ˙Ilker Kuroz, 1994; Hajic and Hladka, 1998; Mueller et al., 2013). Following (M¨uller and Schuetze, 2015), we also add the part-of-speech to this morphological tag and refer to it as POS-MORPH: I see four words | POS=noun:CASE=acc:· · · · · · :NUMBER=plural Given a word in context, we predict a POS- MORPH tag as a complete unit, rather than as the individual component parts. This approach al- lows us to share large parts of the model but can only produce POS-MORPH analyses attested in the training data (cf. Table 2). This is still the stan- dard approach to morphological tagging and dis- ambiguation as, given sufficient amounts of train- ing data, the number of POS-MORPH descrip- tions that cannot be produced usually is …

Content truncated to 500 words