printer friendly version of this page

Processing Morphologically-Rich Languages

Tutorial at INTERSPEECH 2007, Antwerp, Belgium

Morphologically-rich languages like Arabic, Turkish, Finnish, Korean, etc., present significant challenges for speech processing, natural language processing and machine translation. These languages are characterized by highly productive morphological processes (inflection, agglutination, compounding) that may produce a very large number of word forms for a given root form.  Modelling each form as a separate word leads to a number of problems for speech and language processing applications, including:

1. increase in dictionary size: the size of word lists and dictionaries, such as pronunciation dictionaries in speech recognizers, explodes, which places heavy demands on computational resources, such as memory and processing time of speech decoders
2. poor language model (LM) probability estimation: many combinations of word forms will have been observed rarely or not at all in language model training data, resulting in unreliable probability estimates
3. higher out-of-vocabulary (OOV) rate:  when moving from a known (training) set to an unknown (test) set, many novel word forms will occur that need to be accounted for
4. inflection gap for machine translation:  multiple different forms of  the same underlying baseform are often treated as unrelated items, with negative effects on word alignment and translation accuracy.

Large-scale speech and language processing systems require more advanced modelling techniques to address these problems:

• automatic decomposition of complex word forms into smaller units
• methods for optimizing the selection of units at different levels of processing
• diacritization/vowelization  (for Arabic)
• pronunciation modelling for morphologically-rich languages
• morphologically-rich languages in speech synthesis
• novel probability estimation techniques that avoid data sparseness problems
• creating data resources and annotation tools for morphologically-rich languages

Presenters

Katrin Kirchhoff
EE Department, University of Washington

Ruhi Sarikaya
IBM T.J. Watson Research Center

Short Bios

Dr. Kirchhoff is a Research Assistant Professor in the EE Department at the University of Washington. Her research interests are in automatic speech recognition and natural language processing, with an emphasis on multilingual applications.  She has published over 50 refereed conference papers, journal papers, and book chapters in these areas and is co-editor of a recent book on “Multilingual Speech Processing”. In 2002 she led a team effort on developing Novel Models for Arabic Speech Processing at the Johns-Hopkins Summer Research Workshop.  Dr. Kirchhoff has served on numerous conference and workshop committees and is a member of the Editorial Board of the Speech Communication journal.