French relative clauses introduced by dont (from now on: dont-RC) have been well investigated on the theoretical level (Godard 1988) but have not lead to many empirical studies (Gapany 2004). Blanche-Benveniste (1995) observed that dont-RC are not very common in spoken French and restricted to the extraction of a few verb complements. The aim of this article is to describe dont-RC on the basis of three empirical studies: two corpus studies and one online experimental study. The corpus studies have been conducted on a written corpus (French Treebank) and on a spoken corpus (Corpus de Français Parlé Parisien des années 2000). The results confirm partly Blanche-Benveniste's claim: the written usage of dont-RC differs from its spoken use, but it is not restricted in the later to specific idioms. Both corpora corroborate the work of Godard (1980) : dont is widely used as a complement on noun, extractions out of NP are not marginal, not even out of subject NP which are considered non-grammatical by some linguistic theories. The corpus studies show that this kind of extraction is widely-used in French and even concern “real subjects” of transitive verbs, which goes against the prediction made by many generative theories. In generative grammar, relative clauses (as well as any extraction) are subject to island constraints (Ross 1957). The nominal island constraint predicts that the extraction of a noun complement should be harder than the extraction of a verb complement. The subject island constraint rules out extractions out of subject NPs, which is claimed to be only acceptable when the subject NP is an underlying object (Chomsky 2008). However, in our corpora, dont-RC extractions out of NP subject are more frequent than extractions out of NP object. This observation confirms Gibson's Dependency Locality Theory (DLT), which predicts that the extraction from subject NP should have less cognitive costs than the extraction from object NP. Other observations corroborate this theory as well: for example, that dont-RC corresponding to the extraction of the complement of a verb favor pronominal or inverted subjects. Our experimental study is an online acceptability task. Its aim is to compare the judgments from native speakers related to dont-RC implying an extraction out of subject NP and object NP (the kind of subject involved in the later case has an impact of the DLT's predictions and have been tested as well). The results confirm again the predictions from the DLT: the judgments for the extraction out of subject NPs are significantly higher than for the extraction out of object NPs.
The results of the online study can be downloaded here: