Extending the Use of Adaptor Grammars for Unsupervised Morphological Segmentation of Unseen Languages

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 900–910, Osaka, Japan, December 11-17 2016. Extending the Use of Adaptor Grammars for Unsupervised Morphological Segmentation of Unseen Languages Ramy Eskander∗ Owen Rambow∗ ∗ Center for Computational Learning Systems †Department of Computer Science Columbia University New York, NY, USA {rnd2110@,rambow@ccls.,ty2313}columbia.edu Tianchun Yang† Abstract We investigate using Adaptor Grammars for unsupervised morphological segmentation. Using six development languages, we investigate in detail different grammars, the use of morphological knowledge from outside sources, and the use of a cascaded architecture. Using cross-validation on our development languages, we propose a system which is language-independent. We show that it outperforms two state-of-the-art systems on 5 out of 6 languages. 1 Introduction Morphological segmentation, the splitting of words into smaller units (morphs), is an important sub-task in several natural language processing (NLP) applications. With the increasing interest in NLP for low- resource languages, unsupervised morphological segmentation becomes a crucial pre-processing step to reduce data sparseness: instead of working on a large vocabulary of plausible words, a smaller set of smaller word units is processed. A well-known toolkit used for unsupervised segmentation is Morfessor (Creutz and Lagus, 2007), which is a generative probabilistic model. In competition, Adaptor Grammars (AGs) (Johnson et al., 2007) represent a framework for specifying compositional nonparametric Bayesian models and are ap- plied in unsupervised segmentation with notable success (Johnson, 2008). AGs generalize probabilistic context-free grammars by allowing some nonterminals to be “adapted, which allows for dependencies between applications of these rules. Sirts and Goldwater (2013) present an in-depth investigation of the use of AGs. They make two important contributions. First, they discuss the effect of the underlying grammar on the results of unsupervised morphological segmentation. Sec- ond, they investigate two ways of using a small amount of annotated data during training. They show that while for English, Morfessor remains the top performing system, on three other languages their approach can beat the high Morfessor baseline. A typical application for unsupervised morphological segmentation involves situations in which we are confronted with a low-resource language for which no prior NLP work exists, and for which we cannot annotate even a small corpus (either for lack of time, or because no annotators are available). Therefore, in this paper, we are interested in remaining language-independent and entirely unsupervised. We make the following contributions: • We follow the insight of Sirts and Goldwater (2013) that the underlying grammar in an unsupervised AG approach matters. We explore a much larger set of grammars. We show that for most languages we develop on, the grammars we propose in this paper allow for the best AG results to date. The grammars are discussed in Section 4. • We explore the use of “scholar-seeded knowledge”. Here, instead of annotating even a small set with the desired result of the machine learning process (a segmentation), we search the web for easily accessible information about affixes in our language of interest, and explicitly include these in the grammar used in …

Content truncated to 500 words