Throughout my career I have been working in several areas of speech processing, natural language understanding and machine learning. Here are some of the most interesting topics:

In the field of language modeling I conducted studies showing advantages of discriminative parameter estimation techniques such as MCE (Minimum Classification Error) and especially MMI (Maximum Mutual Information) for dialog act classification, but also some applications in bioinformatics.

A considerable amount of my attention was dedicated to investigation of potential of prosodical information for various tasks of speech processing, such as emotion and dialict classification but also spotting of alcohol intoxication. Here and throughout my later career it was of particular interest to me to develop algorithms that do not depend on upstream word recognition and thus avoid unneccessary complications due to difficult ASR conditions.

This motif has persisted in my contributions to the "How May I Help You?" project. In the framework of this project I was working on calltype classification as well as named entity processing which I viewed as three (though tightly coupled) subtasks, namely: detection, localization and value extraction. In particular, I was investigating the difficult case of missing in-domain transcriptions to estimate language models that would produce ASR output these algorithms would then be running for. Unsupervised language model adoptation as well as phoneme-based NLU were both found to be promising solutions for this case.

Automated prediction of landslides using geological and geospatial information has obviously nothing to do with speech and language, but it is an interestinig machine learning task per se, and I was gratified to see how this problem can be successfully tackled using familiar methods.

At MIT Media Lab the goal of my research was to investigate how people use natural language to exchange navigational instructions, and schematize the findings in a way that would allow their automatic understanding and interpretation. The result was a set of Navigational Information Units (NIUs): such as moves, turns, positions and orientations - whose interpretations can be decomposed into a number of orthogonal constituents (e.g. move type, reference objects, distance etc.) that one can extract independently out of verbal commands. I was able to show that in the map navigation scenario, combination of individual NIUs by means of dynamic programming paradigm leads to faithful replica of entire paths.

Lately, the scope of my interests was extended into the domains of information extraction and automatic question answering with their applications to ACE and GALE projects.

Worth mentioning are also a few other topics like automatic extraction of probabilistic grammars from collapsed and regular texts and also evaluation of classification systems operating with soft reference labels. Ask me about that if would like to learn more.