September 26, 2013

Enis Rıfat Sert, M.S. 2013

Current position: Software Engineer, Facebook. (email).
M.S. Thesis: Word Context and Token Representations from Paradigmatic Relations and Their Application to Part-of-Speech Induction. Koç University, Department of Computer Engineering. September, 2013. (PDF, Presentation).
Publications: bibtex.php




Abstract:
Representation of words as dense real vectors in the Euclidean space provides an intuitive definition of relatedness in terms of the distance or the angle between one another. Regions occupied by these word representations reveal syntactic and semantic traits of the words. On top of that, word representations can be incorporated in other natural language processing algorithms as features.

In this thesis, we generate word representations in an unsupervised manner by utilizing paradigmatic relations which are concerned with substitutability of words. We employ an Euclidean embedding algorithm (S-CODE) to generate word context and word token representations from the substitute word distributions, in addition to word type representations. Word context and word token representations are capable of handling syntactic category ambiguities of word types because they are not restricted to a single representation for each word type.

We apply the word type, word context and word token representations to the part-of-speech induction problem by clustering the representations with k-means algorithm and obtain type and token based part-of-speech induction for Wall Street Journal section of Penn Treebank with 45 gold-standard tags. To the best of our knowledge, these part-of-speech induction results are the state-of-the-art for both type based and token based part-of-speech induction with Many-To-One mapping accuracies of 0.8025 and 0.8039, respectively. We also introduce a measure of ambiguity, Gold-standard-tag Perplexity, which we use to show that our token based part-of-speech induction is indeed successful at inducing part-of-speech categories of ambiguous word types.

Full post...