Skip to main content

Linguistic Analysis of the European Languages

Objective

The LING-ANALYSIS project produced the software necessary to perform grapheme-to-phoneme and phoneme-to-grapheme conversion at word level. This involved conversions between the textual and acoustical representation of words and the acquisition of the knowledge required to include speech in the man-machine interface. A linguistic model, based on typical syntactic patterns extracted from texts by statistical analyses, has been developed to deal with ambiguous solutions. The project covered the following languages: Dutch, English, French, German, Greek, Italian and Spanish. The first step was the development of a common methodology among the different languages in order to provide coherent and comparable results. Hardware and software tools were standardisedamong the partners, and language-specific tools developed where necessary. Reference corpora of about 200000 words plus dictionaries and lists of ambiguities (homographs and homophones) were extracted from common European Community texts and newspapers.The efinition and development of a linguistic model for the semi-automatic labelling of new text corpora and for phoneme-to-grapheme conversion on the basis of a contextual analysis was achieved. ed.
The project produced the software necessary to perform grapheme to phoneme and phoneme to grapheme conversion at word level. This involved conversions between the textual and acoustical representation of words and the acquisition of the knowledge required to include speech in the man machine interface. A linguistic model, based on typical syntactic patterns extracted from texts by statistical analyses, has been developed to deal with ambiguous solutions. The project covered the following languages: Dutch, English, French, German, Greek, Italian and Spanish. The first step was the development of a common methodology among the different languages in order to provide coherent and comparable results. Hardware and software tools were standardized among the partners, and language specific tools developed where necessary. Reference corpora of about 200 000 words plus dictionaries and lists of ambiguities (homographs and homophones) were extracted from common European Community texts and newspapers. The definition and development of a linguistic model for the semiautomatic labelling of new text corpora and for phoneme to grapheme conversion on the basis of a contextual analysis was achieved.
The following results are now available for the different languages:
Conversion Algorithm: word level grapheme-to-phoneme and phoneme-to-grapheme conversion algorithms.
Analysis of Language at Word Level
-computer-readable common phonemic alphabet
-consistent systems of grammatical classes
-labelling of text corpora of a few thousand words
-dictionaries, extracted from the corpora, providing (for each word) graphemic and phonemic representations, possible grammatical tabs, and usage frequency
-statistics, extracted from the dictionaries, providing: phonemes and phoneme cluster frequency; graphemes and grapheme cluster frequency; word distribution based on the grapheme length and on the length with or without frequency weighting; and the set f unction K(n) providing K, the percentage coverage of the corpora obtained with the n most frequent words.
Disambiguation Rules for Phoneme-to-Grapheme Conversion
-list of ambiguous words and ambiguity frequency estimates regarding the grapheme/phoneme/grapheme conversions
-transition matrices providing the observed frequency of any pair or triplet of grammatical classes.
Assessment of Conversions: methodologies for evaluating the statistical validity of the information appearing in the transition matrices and for comparing the expected performance in speech recognition of different class systems.
Integration in a Practical Conversion System: a blackboard model of the language that uses the available knowledge on contextual constraints for solving the ambiguities consequent to the phoneme to grapheme conversion and for selecting the most likely sentence from a word lattice.
Exploitation
Full industrial exploitation of the results is expected in the early 1990s in speech processing based systems. Target application areas are unrestricted texts, speech synthesis, and large vocabulary speech recognition. The acquired knowledge and the results obtained will also be useful for applications in other domains, such as optical scanning, word-processing and automatic translation.

Coordinator

Ingegneria C. Olivetti and C. SpA
Address
Corso Svizzera 185
10149 Torino
Italy

Participants (7)

Acorn Computers Ltd
United Kingdom
Address
Acorn House Vision Park Histon
CB4 4AE Cambridge
Centre National de la Recherche Scientifique (CNRS)
France
Address

91406 Orsay
KATHOLIEKE UNIVERSITEIT NIJMEGEN
Netherlands
Address
Erasmusplein 1
6525 HT Nijmegen
RUHR-UNIVERSITY BOCHUM
Germany
Address
Universitätsstraße 150
44780 Bochum
Tecnopolis Csata Novus Ortus
Italy
Address
Strada Provinciale Per Casamassima Km 3.00
70010 Valenzano Bari
UNIV NACIONAL DE EDUCACION A DISTANCIA
Spain
Address
Pabellon Gobierno
X Madrid
UNIV OF PATRAS
Greece
Address

26500 Patrai