Text classification and information extraction

Since I started at Kofax, the main part of my work shifted to general document processing. We are mainly involved in the classification of documents and the extraction of information from them. Interestingly, some aspects combine very well with finite state approaches.

Natural Speech Processing

The goal of Natural Speech Processing is to employ speech recognition and processing systems that facilitate capabilities close to those of humans. Some main properties of such systems are

• Incrementality. Only if a speech system begins processing its input before the speaker has finished her utterance, a natural, fluid reaction is possible; for example, a machine is only able to interrupt a dialog partner if it understands while it is listening.
• Interactivity. The use of top-down interactions, the process of influencing understanding phases using knowledge from higher levels, is crucial in several ways for natural speech understanding. Ambiguity that arises in certain components can be reduced based on evidence from other modules; the search spaces within a system can be modified, leading to an expectation-based understanding as is often used by people.
• Parallelism. The efficiency of computers is steadily growing. However, computers never seem fast enough to cope with the extremely high demands of full speech and language processing. Parallelism (both on an intra-modular and an inter-modular level) can alleviate some of the deficiencies. Moreover, only a modular, incremental, parallel system can truly exploit interaction, and thus be a part of a Natural Speech Understanding system

Some of my work has been in the vicinity of some of these properties. My master's thesis dealt with parallel parsing on a transputer network (Available here).

My dissertation describes a fully incremental, parallel system for translating spontaneous speech, showing that incremental techniques are indeed ready to be employed in next-generation speech processing systems. An English version is published by Springer in the LNAI series: Jan W. Amtrup (1999). Incremental Speech Translation. Springer Verlag, Berlin, Heidelberg, New York. 1999. Number 1735 in Lecture Notes in Artificial Intelligence. Springer offers online access to the book at http://www.springerlink.com/content/f1m9jc9135jx/.

Persian Morphology and Machine Translation

At CRL, I worked in the Shiraz Project, which created a machine translation system from Persian to English. We used finite state morphology with feature structures, feeding into a typed unification-based syntactic parser. Over time, I developed the morphology into a two-level system using finite state transducers with typed feature structures as weights on the edges and proved that having feature structures does not come with a price in efficiency. A short description of the architecture is here.

Architecture of NLP systems

Starting with Verbmobil, and continuing through my time at CRL and Bowne Global, I was interested in the architecture of natural language processing systems, in particular of machine translation systems. This extended both to the makeup of components and their interaction and to the general software architecture.