The Zodiac project
Zodiac is a system that will automatically restore accents (diacritical marks) in a French text. It supersedes an older technology, Réacc, and benefits from recent technological developments.
To obtain a license for Zodiac or for any question or comment on Zodiac, please contact Guy Lapalme.
You can test Zodiac right here.
A few examples
Here are a few example sentences to try Zodiac with. The sentences do not contain accents or other diacritical marks.
- J'ai ete tres epate par ce systeme car il a une capacite interessante a distinguer les differentes possibilites d'un meme mot en francais.
- Les eleves sont bien eleves car le maitre les eleve bien.
- La cote francaise est bien jolie, mais de son cote, la cote des francais est tres elevee.
- La peche etant son peche mignon, il a peche a la riviere en mangeant une peche.
- Nous avons tache de developper un systeme sans tache.
You can create other examples using this list of French words particularly ambiguous without their accents and diacritical marks.
Zodiac in Microsoft Word
It is possible to use the software within Microsoft Word, on Windows.
» See the Word add-in installation page.
Zodiac on Mac
It is possible to use Zodiac as a service and a command-line utility on Mac.
» See the Word service installation page.
People who exchange e-mail in French are often confronted, even today, with text containing characters that should be accented but aren't. There are two explanations for this phenomenon. Firstly, for many years, there was no widely accepted standard for encoding accented characters within the computer industry, and this resulted in a host of problems when it came to transmitting and processing French text. Secondly, keying-in accented characters can still be a difficult, sometimes acrobatic task, owing as much to ergonomic considerations as to the question of standards. The upshot of all this is that many French speakers systematically avoid using accented characters, at least in their e-mail.
If this situation remains tolerable in practice, it is only because the disappearance of accents does not usually render a French text incomprehensible for a human reader. At worst, it is an irritant that may slow down the reading of the text, in the same way that a text written entirely in upper case letters will.
While French without accents may be acceptable in certain situations, it is certainly not acceptable in general usage, particularly in the case of documents that are to be widely distributed. What is more, the absence of accents can create serious problems for automatic text processing. Whether it be for document retrieval, grammatical, style or spell checking, machine translation, natural language interfaces or any other application, French texts would generally be expected to include accents. This accounts for much of the recent interest in methods for automatically recovering accents, or automatic reaccentuation.
We have developed a tool that automatically inserts accents and other diacritical markings in a text that has been stripped of them. On French texts, our Zodiac system achieves a high level of accuracy, using a probabilistic language model.
How does it work?
The addition of French diacritical marks (accents, umlauts, cedillas) is a two-step process. First, for each word in the text, the program produces a list of accentuation hypotheses. For instance, for the word form
cote, il will produce the list
[cote, coté, côte, côté]. For this, Zodiac queries a French dictionary.
The second step is disambiguation: The program picks the most appropriate form from each list. This choice is guided by a probabilistic language model. This type of model allows the program to pick the most probable form in a given context. Therefore, Zodiac does not base its choice on a set of grammar rules, but rather on a statistical knowledge of language, a knowledge it has acquired by automatically analyzing great quantities of text.
Naturally, the system makes mistakes sometimes. However, we carried out performance tests that have shown that these mistakes occur on average every 200 words, that is to say 2 or 3 accentuation errors per page.
Project manager: Guy Lapalme