TY  - THES
T1  - Computational modeling of lexical ambiguity
A1  - Li,Linlin
Y1  - 2012/11/02
N2  - Lexical ambiguity is a frequent phenomenon that can occur not only for words but also on the phrase level. Natural language processing systems need to efficiently deal with these ambiguities in various tasks, however, we often encounter such system failures in real applications. This thesis  studies several complex phenomena related to word/phrase ambiguity at the level of text and proposes computational models to tackle these phenomena. Throughout the thesis, we address a number of lexical ambiguity phenomena varying across the sense granularity line. We start with the idiom detection task, in which candidate senses are constrained toliteral' and idiomatic'. Then, we move on to the more general case of detecting figurative expressions. In this task,  target phrases are not lexicalized but rather bear nonliteral semantic meanings. Similar to the idiom task, this one has two candidate sense categories (literal' and nonliteral').  Next,  we consider a more complicated situation where words often have more than two candidate senses and the sense boundaries are fuzzier, namely word sense disambiguation (WSD). Finally, we discuss another lexical ambiguity problem in which the sense inventory is not explicitly specified, word sense induction (WSI).Computationally, we propose novel models that outperform state-of-the-art systems. We start with a supervised model in which we study a number of semantic relatedness features combined with linguistically informed features such as local/global context, part-of-speech tags,  syntactic structure, named entities and sentence markers. While experimental results show that  the supervised model can effectively detect idiomatic expressions, we further improve the work by proposing an unsupervised bootstrapping model which does not rely on human annotated data but  performs at a comparative level to the supervised model.  Moving on to accommodate other lexical ambiguity phenomena, we propose a Gaussian Mixture Model that can be used not only for detecting idiomatic expressions but also for extracting  unlexicalized figurative expressions from raw corpora automatically.  Aiming at modeling multiple sense disambiguation tasks within a uniform framework, we propose a probabilistic model (topic model), which encodes human knowledge as sense priors via paraphrases of gold-standard sense inventories, to effectively perform on the idiom task as well as two WSD tasks. Dealing with WSI, we find state-of-the-art WSI research is hindered by the deficiencies of evaluation measures that are in favor of either very fine-grained  or very coarse-grained cluster output. We argue that the information theoretic V-Measure is a promising approach to pursue in the future but should be based on more precise entropy estimators, supported by evidence from the entropy bias analysis, simulation experiments, and stochastic predictions.  We evaluate all our proposed models against state-of-the-art systems on  standard test data sets, and we show that our approaches advance the state-of-the-art.
KW  - Ambiguität
KW  - Phraseologie
KW  -  Computerlinguistik
KW  - Maschinelles Lernen
CY  - Saarbrücken
PB  - Universitäts- und Landesbibliothek
AD  - Postfach 151141, 66041 Saarbrücken
UR  - http://scidok.sulb.uni-saarland.de/volltexte/2012/4973
ER  -