Vera Demberg - Research Interests
I'm interested in language processing in humans, computational models of human language processing, parsing (in particular incremental parsing) and dialogue systems.
Education and Research
I'm currently heading a Junior Research Group at the MMCI, Saarland University, Germany.During my PhD at the Institute for Communicating and Collaborative Systems (ICCS), University of Edinburgh I worked under the supervision of Frank Keller and Fernanda Ferreira. I'm interested in using statistical methods from computational linguistics for modelling human sentence processing. As part of my PhD, I developed a psycholinguistically motivated version of tree-adjoining grammar, PLTAG, and implemented the first strictly incremental, fully connected parser. In spring 2008, I spent three month working on a spoken dialog system during my internship with Google / Mountain View.
From 2001 to 2006, I studied computational linguistics with computer science as a minor at the Institute for Natural Language Processing (IMS), University of Stuttgart. My thesis, which I did with IBM Research and Development GmbH, Böblingen, was concerned with letter-to-phoneme conversion for the German IBM Text-to-Speech Synthesis System. I investigated the benefit of morphological preprocessing for German speech synthesis, compared different approaches to syllabification, stress assignment and grapheme-to-phoneme (g2p) conversion, and implemented a joint n-gram model for g2p that incorporates language-independent phonological constraints which help to boost performance significantly.
I'm also interested in unsupervised methods for morphological segmentation (see Software and Publications).
In 2004 / 2005, I obtained an MSc in Artificial Intelligence from the University of Edinburgh. The topic of my Master's thesis was "Information Presentation in Spoken Dialogue Systems". I developed a strategy for presenting a large number of options in spoken dialogue systems and implemented the approach in a flight recommendation system. This system exploits a user-model to identify the options that are particularly compelling to a given user. Multiple attractive options are structured and their respective advantages and disadvantages are compared explicitly.