Yuancheng Tu
10103 Bravern 1
Bellevue, WA, 98004
425-7073110
myFirstName_myLastName (at) hotmail dot com

I recently graduated and moved to Microsoft Speech Recognition and Synthesis team. Before that, I was a Ph.D student at the University of Illinois at Urbana-Champaign, majoring in computational linguistics. I worked in the Cognitive Computation Group and my thesis advisor is Dan Roth. My primary research interests are natural language processing, machine learning, computational lexical semantics. I am also interested in structure learning in NLP and Text Mining.

DSSI-2011 Yahoo-Award Winning Project: Expert Search by Non-Experts

We propose to develop an expert and expertise search system which tracks and retrieves researchers and professors based on their research expertise, visualizes their related information via an automatic generated webpages. Our system will have many useful implications for the academic, commercial and government fields. For example, academic funding agencies frequently search for researchers which are not only experts in their fields but also satisfy from geographical constraints. Commercial implications include, but are not limited to, an email agent that could prompt people of events when and only when the person is likely to be interested in that event. Government implications include military intelligence gathering, in that, a ranking intelligence officer may want to communicate with an expert on a topic who is nearby.

This project introduces a novel unsupervised sequence learning model for identifying English Multiword Expressions (MWEs). We model the compositionality of MWEs in a given context as a sequence learning problem over each component of the given MWE. When idiomatic uses of a MWE are indistinguishable on the surface from its literal uses in a given context, our model makes a prediction based on the learned best hidden sequence combination and identifies which component can be simplified by examining the binary value of each hidden state. This model proposes a unified way to deal with all types of MWEs in a computational system and provides less ambiguous lexical representation for further knowledge acquisition and inference in other applications such as machine translation, information extraction and data mining.

Lexical Simplification in Textual Inference

In this research we propose to learn the natural language entailment relation via exploiting the logic entailment or polarity expressed by natural language lexicon, such as factive verbs and imperative verbs. The model is rule-based and can be used to generate the entailed statement for a given sentence with a factive or imperative verb as the predicate. This model can be used in textual entailment to produce less ambiguous lexical representation and it can also help to achieve better recall in information retrieval or data mining system.

Some of my previous projects