Course in Weighted Finite-State Transducers in Text and Speech Processing (1 - 2 cr)

Lecturer

Mehryar Mohri (AT&T Labs - Research)

Abstract

Weighted finite-state transducers are used in many text and speech processing applications. They are now used in a variety of applications such as computational biology and speech and image processing.

In this series of lectures, we will describe and analyze some of the fundamental weighted automata and transducer algorithms including the composition, epsilon-removal, determinization, minimization, and synchronization of weighted automata.

Many of these algorithms have been incorporated in the AT&T finite-state machine library (AT&T FSM library) and the grammar library (GRM library). The FSM and GRM libraries are general-purpose software tools for creating, combining, optimizing, and searching weighted finite-state transducers and weighted grammars. We also give a brief introduction to these libraries.

Finally, we will present some recent applications of weighted transducers to speech recognition, including the representation of each component of a recognizer, the efficient construction of an integrated recognition transducer, and the substantial speed-up due to the use of a general weight-pushing algorithm.

Location

Department of General Linguistics
Siltavuorenpenger 20 A
Helsinki

Duration and Organization

The course will be organized in the following way with the following hours for lectures (L) and practical course (P):

May 12, 2003	May 13, 2003	May 14, 2003	May 15, 2003
09:00-12:00 2L + 1P	09:00-12:00 2L + 1P	09:00-12:00 2L + 1P	09:00-12:00 2L + 1P
01:00-04:00 2L + 1P	01:00-04:00 2L + 1P	01:00-04:00 2L + 1P

Prerequisites

The skills required are basic familiarity with computer science algorithms and representations.

Articles

Participants are expected to download these articles and study them before the beginning of the course:

• Mehryar Mohri. Finite-State Transducers in Language and Speech Processing. Computational Linguistics, 23:2, 1997. http://citeseer.nj.nec.com/14002.html
• Mehryar Mohri, Fernando C. N. Pereira, and Michael Riley. Weighted Finite-State Transducers in Speech Recognition. Computer Speech and Language, 16(1):69-88, 2002. http://www.research.att.com/~mohri/postscript/csl01.ps
• Mehryar Mohri. Weighted Grammar Tools: the GRM Library. In Jean Claude Junqua and Gertjan van Noord, editors, Robustness in Language and Speech Technology. pages 165-186. Kluwer Academic Publishers, The Netherlands, 2001. http://www.research.att.com/~mohri/postscript/grm.ps

Do you find these articles difficult? Check out the helper for the background material of the course written by Anssi Yli-Jyrä. Note:Anssi will give an introductory lecture on the subject on 28 April 2002 from 14:00 o'clock at Siltavuorenpenger. Join in if you are interested! There will be lots of time for questions.