Coling 2008

Manchester, 18-22 August, 2008

The 22nd International Conference on Computational Linguistics

Home · Programme · Info for authors · Workshops · Tutorials · Registration · Venue · Accommodation · Sponsorship · Archive · ICCL

Tutorial T4
Regulus: Spoken Language System Construction Without the Agonizing Pain

Sunday August 17, 2.00 - 5.30

Outline· Structure· Prerequisites· Instructors

Outline

Many computational linguists are in principle interested in experimenting with systems that can process spoken language, but feel daunted by the practical problems involved. It is common to hear versions of all of the following objections:

Using spoken language requires a mastery of arcane technologies and large amounts of C++ programming;

Costly data collection exercises are required in order to acquire data to train speech recognisers, and the resulting systems still perform poorly;

Researchers primarily interested in other components of a spoken language system want to put their effort there rather than in building speech recognisers.

In this tutorial, we will present an overview of Regulus, an Open Source toolkit for building grammar-based speech-enabled systems, that is explicitly designed to address these issues. Regulus sits on top of the Nuance platform, and provides tools which support compilation of linguistically motivated feature grammars into Nuance-compatible speech recognisers. Small feature grammars can be written by hand. More substantial grammars are usually based on existing Regulus resource grammars, which can be automatically specialised into efficient domain-specific phrasal grammars using Regulus tools. The specialisation is driven by domain corpora, which initially can be as small as 50 to 100 examples; as corpus data accumulates, coverage of the domain-specific grammar improves. The same data can also be fed into underlying Nuance utilities to perform statistical tuning of the generated recogniser. The toolkit also includes other compilers, which can be used to transform Regulus grammars into normal parsers and generators.

Regulus provides a integrated development environment which allows development of grammars and grammar-based applications in both text and speech mode. Normal working practise is that the developer will first debug her grammars and other rule-sets in text mode. At any point, she can compile the grammar into a recogniser, and test the resulting live speech system without leaving the development environment. This ability to switch seamlessly between text and speech views of the system greatly simplifies the development cycle.

The development environment contains extensive support both for spoken dialogue applications, based on a version of information state update semantics, and also for interlingua-centered speech translation. It also contains tools for constructing embedded help systems, driven by back-up statistical recognisers, which give users active feedback on the grammar's limitations, and lead them into its coverage. Evaluations we have carried out show that grammar-based applications equipped with help facilities of this kind are generally perceived as user-friendly and easy to learn.

The tutorial will cover enough ground that attendees should then be able to download material from the Regulus website and use it to build simple speech recognisers, spoken dialogue systems and speech translators. Regulus makes it easy to build toy systems, and is thus very suitable for teaching purposes; we have already used it as the core technology for courses at UC Santa Cruz and Geneva University. It is also appropriate for serious language engineering, and has been used to construct substantial speech-enabled applications. These include NASA's Clarissa, which has been tested on the International Space Station, and Geneva University's MedSLT, a multi-lingual speech translator for doctor-patient examination dialogues.

Regulus is described in detail in our book Putting Linguistics into Speech Recognition (CSLI Press, 2006) which received positive reviews from Computational Linguistics and the Journal of Natural Language Engineering.
Approximately two thirds of the tutorial consists of material covered in the book; the other third presents new work.

Structure

The tutorial will be structured as follows:

Introduction and overview (20 minutes). What grammar-based recognition is; why it can be useful; what the Regulus system does; demo of one or two Regulus-based applications.
Basics of Regulus (1 hour). Compiling feature grammars into CFG form; grammar specialisation using Explanation Based Learning; the Regulus resource grammars; using corpus data to perform statistical tuning; semantic representations; building embedded help systems; knowledge based rescoring of N-best speech hypothesis lists.
Using Regulus resource grammars (20 minutes). Adding domain-specific lexicon entries; using the grammar specialiser; regression testing and the development cycle.
Building spoken dialogue applications (40 minutes). Information state update semantics; logical forms and dialogue moves; context-sensitive processing; asynchronous dialogue architecture; regression testing and the development cycle.
Building translation applications (20 minutes). Compiling Regulus grammars into generators; using Regulus grammar to define interlingua; translation rules; regression testing and the development cycle.

Prerequisites

We will assume some familiarity with feature grammars. A basic knowledge of speech technology will be helpful, but is not required.

Instructors


Beth Ann Hockey UCSC UARC Mail Stop 19-26 NASA Ames Research Center Moffet Field CA 94035, USA Email: bahockey at ucsc.edu	Manny Rayner TIM/ISSCO, University of Geneva 40 bvd du Pont-d'Arve CH-1211 Geneva 4 Switzerland Email: Emmanuel.Rayner at issco.unige.ch

Dr. Beth Ann Hockey is an Adjunct Associate Professor of Linguistics at UC Santa Cruz and a Senior Researcher for the UC Santa Cruz UARC at NASA Ames Research Center. Dr. Hockey is also a core developer for the Open Source Regulus project, and co-author of the book Putting Linguistics into Speech Recognition: The Regulus Grammar Compiler (2006 CSLI Press). Her areas of expertise include speech translation systems, spoken dialogue systems, language modelling, grammar development, embedded help systems, discourse and phonetics. She was the project lead and a principle designer and developer of the Clarissa Procedure Navigator, which in 2005 became the first spoken dialogue system used in space. She is project lead for the UC Santa Cruz - Ford Motors spoken dialogue project, and a developer on the MedSLT medical speech translator project. At UC Santa Cruz, Dr. Hockey teaches a popular course on Spoken Dialogue Systems. She is the author of over 50 refereed publications.

Dr. Manny Rayner has been the lead developer on the Regulus project since its inception, and has held senior positions at SRI International, NASA Ames Research Center, and several speech and language startups. He has worked actively on numerous areas of speech and language technology, including language modelling, speech translation, spoken dialogue systems, machine learning, grammar engineering, logic programming and construction of development environments for speech and language systems. He has over 100 refereed publications, including books on the Spoken Language Translator and Regulus systems.

BACK to Tutorials main page