Intelligent feedback on learners’ full written
sentence productions requires the use of Natural Language Processing (NLP)
tools and, in particular, of a diagnosis system. Most syntactic parsers, on which
grammar checkers are based, are designed to parse grammatical sentences and/or
native speaker productions. They are therefore not necessarily suitable for
language learners. In this paper, we concentrate on the transformation of a
French syntactic parser into a grammar checker geared towards intermediate to
advanced learners of French. Several techniques are envisaged to allow the
parser to handle ill-formed input, including constraint relaxation. By the
very nature of this technique, parsers can generate complete analyses for
ungrammatical sentences. Proper labelling of where the analysis has been able to
proceed thanks to a specific constraint relaxation forms the basis of the error
diagnosis. Parsers with relaxed constraints tend to produce more complete,
although incorrect, analyses for grammatical sentences, and several complete
analyses for ungrammatical sentences. This increased number of analyses per
sentence has one major drawback: it slows down the system and requires more
memory. An experiment was conducted to observe the behaviour of our parser in the
context of constraint relaxation. Three specific constraints, agreement in number,
gender, and person, were selected and relaxed in different combinations. A learner
corpus was parsed with each combination. The evolution of the number of correct
diagnoses and of parsing speed, among other factors, were monitored. We then
evaluated, by comparing the results, whether large scale constraint relaxation is
a viable option to transform our syntactic parser into an efficient grammar
checker for CALL.