Extraction of the Multi-word Lexical Units in the Perspective of the Wordnet Expansion

Proceedings of Recent Advances in Natural Language Processing, pages 512–520, Hissar, Bulgaria, Sep 7–9 2015. Extraction of the Multiword Lexical Units in the Perspective of the Wordnet Expansion Maciej Piasecki Wrocław Univ. of Technology maciej.piasecki@pwr.edu.pl Michał Wendelberger Wrocław Univ. of Technology michal.wendel@gmail.com Marek Maziarz Wrocław Univ. of Technology mawroc@gmail.com Abstract The paper focuses on selecting an optimal set of the Multiword Expressions Extrac- tion methods used as a tool during word- net expansion. Wordnet multiword lexi- cal units are a broad class and it is dif- ficult to find a single extraction method fulfilling the task. Many extraction as- sociation measures were tested on very large corpora and a very large wordnet, namely plWordNet. Several new measures are proposed and compared with selected methods in the literature. Two ways of combining measures into ensembles were analysed too. We showed that method se- lection and the tuning of their parameters can be transferred between two large cor- pora. The comparison of the extracted col- locations with the huge set of plWordNet multiword lexical units revealed that the performance of the methods is much be- low the optimistic levels reported in the literature. However, the carefully selected set and combination of the methods can be a valuable tool for lexicographers. 1 Introduction A large number of different methods for the extraction Multiword Expressions (henceforth, MWE) have been proposed in literature. Most of them are focused on particular properties of MWEs, e.g. non-compositionality. Before apply- ing selected methods as the support for the con- struction of a large lexicon we have to answer sev- eral questions. • What method should we select if we need to extract MWEs of different subtypes? • How effective are different methods in help- ing lexicographers who use a complex but well specified definition of Multiword Lexi- cal Units (MLUs)1 (Maziarz et al., 2015a)? • What is the performance of the known ex- traction methods when they are applied to big corpora (e.g. >1 billion words) and next eval- uated against very large lexicons of MWEs? We aim at the development of a method for the extraction of MLUs from large corpora for the needs of wordnet expansion. MLUs encom- pass a broad spectrum of MWEs: from non- compositional MWEs to specialist terminology. The starting point for this work were the seminal papers of Peˇcina, e.g. (Peˇcina, 2010), including tests of a very large number of MWE extraction methods. However, those tests were done on rela- tively limited data set. The corpus used by Peˇcina in his experiments consisted of about 1.5 million words. In our work we utilised different corpora including the testing Merged Corpus of 1.6 billion words covering rich variety of topics and genres2. So, it was more than 1 000 times bigger than that used in (Peˇcina, 2010). We wanted to perform large scale evaluation done on big corpora, utilising a very large lexi- con of MWEs and focused on Polish, a language which is significantly different from English. The first experiments (completed) inspired …

Content truncated to 500 words