Weakly Supervised Morphology Learning for Agglutinating Languages Using Small Training Sets.

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 976–983, Beijing, August 2010 Weakly Supervised Morphology Learning for Agglutinating Languages Using Small Training Sets Ksenia Shalonova Computer Science, University of Bristol ksenia@cs.bris.ac.uk Bruno Gol´enia Computer Science, University of Bristol csbsgg@bristol.ac.uk Abstract The paper describes a weakly supervised approach for decomposing words into all morphemes: stems, prefixes and suffixes, using wordforms with marked stems as training data. As we concentrate on under-resourced languages, the amount of training data is limited and we need some amount of supervision in the form of a small number of wordforms with marked stems. In the first stage we in- troduce a new Supervised Stem Extrac- tion algorithm (SSE). Once stems have been extracted, an improved unsupervised segmentation algorithm GBUMS (Graph- Based Unsupervised Morpheme Segmen- tation) is used to segment suffix or prefix sequences into individual suffixes and pre- fixes. The approach, experimentally val- idated on Turkish and isiZulu languages, gives high performance on test data and is comparable to a fully supervised method. 1 Introduction The major function of morphological analysis is decomposition of words into their constituents - stems and prefixes/suffixes. In recent years Ma- chine Learning approaches were used for word de- composition. There is a number of both unsuper- vised morphology learning systems that use ”raw” wordforms as training data (Creutz and Lagus, 2002; Goldsmith, 2001; Kazakov and Manand- har, 2001) and supervised morphology learning systems using segmented wordforms into stems and affixes as training data (Oflazer et al., 2001). The supervised morphology learning systems are usually based on two-level morphology (Kosken- niemmi, 1983). There is also a weakly supervised approach that uses, for example, wordpairs as in- put, and this was applied mainly to fusional lan- guages for stem extraction (Erjavec and Dzeroski, 2004). Our project concerns developing speech technology for under-resourced languages. For this type of applications we need a relatively fast, cheap (i.e. does not require large training sets), almost knowledge-free approach that gives high performance. We have chosen to use wordforms with marked stems as training data in order to ful- fill the criteria mentioned above. Morphological analysis is used in many prac- tical Natural Language Processing applications such as Machine Translation, Text Mining, spell- checkers etc. Our near-term goal is the integration of the morphology learning algorithms into the language-independent Text-to-Speech (TTS) sys- tem for improvement of grapheme-to-phoneme rules, stress prediction and tone assignment. In particular, the morphology learning algorithms described in this paper will be incorporated into the available isiZulu TTS system for automatic prediction of lexical tones. In the isiZulu language lexical tone assignment depends on the morpheme boundary. The current isiZulu TTS system is tone-deaf due to the lack of morphological de- composition. A number of perception tests will be carried out in order to evaluate which perfor- mance of morphology decomposition is accept- able for TTS and will improve the quality of the synthesised speech. It seems that the unsuper- vised morphology learning systems can be rela- tively easy to …

Content truncated to 500 words