An Agglomerative Hierarchical Clustering Algorithm for Labelling Morphs

Proceedings of Recent Advances in Natural Language Processing, pages 129–135, Hissar, Bulgaria, 7-13 September 2013. An Agglomerative Hierarchical Clustering Algorithm for Labelling Morphs Burcu Can Department of Computer Engineering Hacettepe University Beytepe, Ankara 06800 Turkey burcu.can@hacettepe.edu.tr Suresh Manandhar Department of Computer Science University of York Heslington, York, YO10 5GH, UK suresh.manandhar@york.ac.uk Abstract In this paper, we present an agglomera- tive hierarchical clustering algorithm for labelling morphs. The algorithm aims to capture allomorphs and homophonous morphemes for a deeper analysis of seg- mentation results of a morphological seg- mentation system. Most morphological segmentation systems focus only on seg- mentation rather than labelling morphs ac- cording to their roles in words, i.e. inflec- tional (cases, tenses etc.) vs. derivational. Nevertheless, it is helpful to have a better understanding of the roles of morphs in a word to be able to judge the grammatical function of that word in a sentence; i.e. the syntactic category. We believe that a good morph labelling system can also help part- of-speech tagging. The proposed cluster- ing algorithm can capture allomorphs in Turkish successfully. We obtain a recall of 86.34% for Turkish and 84.79% for En- glish. 1 Introduction Most morphological segmentation systems (Creutz and Lagus (2002; Creutz and Lagus (2004; Goldsmith (2001)) perform only the segmentation of words and do not label morphs according to how they function in a word. As a rule, some morphemes function as inflective, whereas some morphemes function as derivative. However, we do not aim to distinguish inflection or derivation within a word, but we aim to dis- tinguish between various types of morphs which are either inflective or derivative, e.g. allomorphs, homophonous morphemes. Labelling morphs not only helps with analysing the segmentation of a word, but can also help other natural language problems, i.e. part-of-speech tagging. The main purpose of this paper is to serve as a post-processing tool to label morphs that have been discovered by a morphological segmenta- tion system. Our main aim is directed towards the Morpho Challenge competition (Mikko Ku- rimo (2011)), which provides a platform to com- pare participant morphological segmentation sys- tems. In Morpho Challenge, morph labels in a seg- mented word and the respective morph labels in its gold standard are compared. Example 1.1 For example, the gold standard analyses of ‘arrangements’ and ‘standardizes’ in Morpho Challenge are given as: arrangements arrange V ment s +PL standardizes standard A ize s +3SG Although in both analyses -s occurs, their la- bels are different; +PL (plural) and +3SG (third person singular). There is not much work done in morpheme la- belling. Spiegler (Spiegler, 2011) presents two algorithms for morpheme labelling: one of them learns morpheme labels once morphological seg- mentation is completed and the other finds mor- pheme labels during morphological segmenta- tion. Both algorithms work in a supervised set- ting in which ground truth morphemes are pro- vided. Bernhard (Bernhard, 2008) suggests an- other morpheme labelling algorithm which labels morphemes as a stem, suffix, base, or prefix. Therefore, the proposed labelling method does …

Content truncated to 500 words