The end of my work on French grammar checking

I have finished my first work on LanguageTool. I have adapted the tool to French grammar checking. The following resume presents the end of this work.
You can download the report ( Mémoire) and the slides ( Soutenance). They are written in French.

Work on rules

As I explained previously, at the beginning of the French grammar checker project, Myriam Lechelt has worked on An Gramadóir. She has written many disambiguation and correction rules. Since An Gramadóir was limited and did not really suit to French, it was abandoned.

During my work, I have converted An Gramadóir's rules to LanguageTool. Thanks to Marcin Miłkowski who implemented a disambiguator, according to my instructions, I could import disambiguation rules as well as correction rules. Moreover, I simplified them a lot and I considerably reduced their number thanks to the XML language.

Then, I have analysed a corpus of mistakes (V. Lucci et A. Millet, 1994, L'orthographe de tous les jours, enquête sur les pratiques orthographiques des français, Editions Champion) and I have extracted  new grammar rules from it.

LanguageTool can detect the following kind of mistakes :

• phonetic proximity (confusion of homophones like ont and on, ça and sa, etc.)

• mistakes in verb phrases (confusion between infinitive and past participle, conjugated form and past participle, etc.)

• subject-verb agreement (personal pronoun or noun phrase with only a determiner and a noun)

Limits of the formalism

While working on the rules, I made tests that showed me the limits of the formalism of LanguageTool. Because of the rigid pattern matching on which it is based, if the patterns described in the rules do not exactly match the text, the rules become inefficient and prevent some mistakes from being detected. Moreover, it is necessary to foresee every wrong combination of words to describe them in the rules. It leads to a combinatory explosion of the number of rules, especially in noun phrases.

The formalism also generates lots of wrong alarms, because of ambiguities or wrong tags. Some mistakes can be detected simultaneously several times by different rules. And when a word is wrong, it can cause wrong alarms on nearby words, since the rules are based on the context.

New formalism

I have developed a new formalism to improve French grammar checking in LanguageTool. It is based on chunks and unification of features structures (see An alternative with chunks and unification). I mix a contextual syntactic theory (chunks, Abney) and a generative syntactic theory (unification, Chomsky). This is not a typical combination, but it makes possible to go further in grammar checking by delimiting an area in the sentence where all words must agree. It is then no longer necessary to describe all wrong combinations of words. Instead of listing agreement mistakes, inconsistencies are detected in phrases.

Conclusion

Thanks to my work for my MPhil, French grammar checking is available for OpenOffice.org. But there is still a lot of work left. It is necessary to create a tool compatible with the new formalism, and to build and analyse a corpus of mistakes to write new grammar rules.

A new approach for grammar checking

To improve grammar checking, I am considering another method which consists in doing at the same time the morphosyntactic analysis and the grammar checking, while the sentence is read. This "left-right" method is based on the principle of latencies (Tesnières, 1959). With the declaration of what is expected after a word or a phrase, inconsistencies can be detected, instead of listing all possible mistakes.
This approach will also solve the problem of the vicious circle in grammar checking. Indeed, for mistakes to be detected, the tagging must not be wrong. But for it to be correct, the text must not contain any mistake...