Semi-Automatic Construction of Word-Formation Networks (for Polish and Spanish)

Semi-Automatic Construction of Word-Formation Networks (for Polish and Spanish) Mateusz Lango1,2, Magda ˇSevˇc´ıkov´a2 and Zdenˇek ˇZabokrtsk´y2 1Poznan University of Technology, Faculty of Computing, Institute of Computing Science 2Charles University, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics mlango@cs.put.edu.pl,{sevcikova,zabokrtsky}@ufal.mff.cuni.cz Abstract The paper presents a semi-automatic method for the construction of derivational networks. The proposed approach applies a sequential pattern mining technique in order to construct useful morphological features in an unsupervised manner. The features take the form of regular expressions and later are used to feed a machine-learned ranking model. The network is constructed by applying resulting model to sort the lists of possible base words and selecting the most probable ones. This approach, besides relatively small training set and a lexicon, does not require any additional language resources such as a list of alternations groups, POS tags etc. The proposed approach is applied to the lexeme sets of two languages, namely Polish and Spanish, which results in the establishment of two novel word-formation networks. Finally, the network constructed for Polish is merged with the derivational connections extracted from the Polish WordNet and those resulting from the derivational rules developed by a linguist, resulting in the biggest word-formation network for that language. The presented approach is general enough to be adopted for other languages. Keywords: derivation, derivational morphology, Polish, Spanish, lexical network, learning to rank, sequential pattern mining 1. Introduction Derivational morphology has moved into focus of Nat- ural Language Processing (NLP) only recently. For some languages, we observe a significant research ef- fort in the construction of resources specialized in derivation, e.g. DerivBase (Zeller et al., 2013) for German, D´emonette (Hathout and Namer, 2014) for French, DerivBase.Hr (ˇSnajder, 2014) for Croatian, Der- iNet (ˇSevˇc´ıkov´a and ˇZabokrtsk´y, 2014; ˇZabokrtsk´y et al., 2016) for Czech, or Word Formation Latin (Litta et al., 2016). However, for many other languages the data re- sources which provide information about derived words are scarce or even lacking. Unfortunately, the creation of such resources requires a considerable human effort and is highly time-consuming. In this paper, we propose a method for semi-automatic con- struction of a derivational network which can be applied to under-resourced languages. The proposed approach re- quires only two resources: a set of lexemes and a relatively small training set which should contain examples of derived lexemes with their base words. First, the method is looking for frequent patterns in a given lexeme set in order to automatically detect character-level regularities in the word construction of the language un- der consideration. The mentioned process is conducted in a completely unsupervised manner and no hand-crafted rules are used. Given those frequent patterns, each lexeme can be described by the presence (or absence) of a given pat- tern. Such lexeme descriptions serve as feature vectors for machine learning techniques and allow to train a ranking model. Later, we start the network construction by making an at- tempt to find a base word for each lexeme. In order to avoid the consideration of all …

Content truncated to 500 words