Supervised Morphological Segmentation in a Low-Resource Learning Setting using Conditional Random Fields

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 29–37, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Supervised Morphological Segmentation in a Low-Resource Learning Setting using Conditional Random Fields Teemu Ruokolainena Oskar Kohonena Sami Virpiojaa Mikko Kurimob a Department of Information and Computer Science, Aalto University b Department of Signal Processing and Acoustics, Aalto University firstname.lastname@aalto.fi Abstract We discuss data-driven morphological segmentation, in which word forms are segmented into morphs, the surface forms of morphemes. Our focus is on a low- resource learning setting, in which only a small amount of annotated word forms are available for model training, while unan- notated word forms are available in abun- dance. The current state-of-art methods 1) exploit both the annotated and unan- notated data in a semi-supervised man- ner, and 2) learn morph lexicons and sub- sequently uncover segmentations by gen- erating the most likely morph sequences. In contrast, we discuss 1) employing only the annotated data in a supervised man- ner, while entirely ignoring the unanno- tated data, and 2) directly learning to pre- dict morph boundaries given their local sub-string contexts instead of learning the morph lexicons. Specifically, we em- ploy conditional random fields, a popular discriminative log-linear model for seg- mentation. We present experiments on two data sets comprising five diverse lan- guages. We show that the fully super- vised boundary prediction approach out- performs the state-of-art semi-supervised morph lexicon approaches on all lan- guages when using the same annotated data sets. 1 Introduction Modern natural language processing (NLP) appli- cations, such as speech recognition, information retrieval and machine translation, perform their tasks using statistical language models. For mor- phologically rich languages, estimation of the lan- guage models is problematic due to the high num- ber of compound words and inflected word forms. A successful means of alleviating this data sparsity problem is to segment words into meaning-bearing sub-word units (Hirsimäki et al., 2006; Creutz et al., 2007; Turunen and Kurimo, 2011). In lin- guistics, the smallest meaning-bearing units of a language are called morphemes and their surface forms morphs. Thus, morphs are natural targets for the segmentation. For most languages, existing resources contain large amounts of raw unannotated text data, only small amounts of manually prepared annotated training data, and no freely available rule-based morphological analyzers. The focus of our work is on performing morphological segmentation in this low-resource scenario. Given this setting, the cur- rent state-of-art methods approach the problem by learning morph lexicons from both annotated and unannotated data using semi-supervised machine learning techniques (Poon et al., 2009; Kohonen et al., 2010). Subsequent to model training, the methods uncover morph boundaries for new word forms by generating their most likely morph se- quences according to the morph lexicons. In contrast to learning morph lexicons (Poon et al., 2009; Kohonen et al., 2010), we study mor- phological segmentation by learning to directly predict morph boundaries based on their local sub- string contexts. Specifically, we apply the linear- chain conditional random field model, …

Content truncated to 500 words