Subject island? (Counter)evidence from French dont-relative clauses

Whether island constraints (Ross 1967) are due to structural (Chomsky 2008, Sprouse et al. 2015, etc.) or cognitive constraints (Klunder 1992, Hofmeister et al. 2013, etc.) is still an open debate. Subject islands have been shown to vary across languages (Rizzi 1990) and across constructions (Sprouse et al. 2015). In French, dont relative clauses (DRC) (Godard 1992) exhibit ‘de’ (‘of which’) complement extraction for verbal (1a), object noun (1b) or subject noun (1c) complements. Mainstream generative approaches predict (1a) to be easier than (1b), and (1b) than (1c). Processing approaches such as DLT (Gibson, 2000) predict the reverse: (1c) should be easier than (1a), and (1a) than (1b), in terms of linear distance between dont and the gap_. Furthermore, DLT predicts (2a), with a 1^st person subject ‘je’, to be easier than (1a) and (2b) than (1b).

1. a. from VP une ville dont mes amis [parlent souvent _ ] (a town of-which my friends often talk )                              

                  b. from object NP une ville dont mes amis apprécient [les commerçants _ ] (a town of-which my friends like the                                          merchants)

                  c. from subject NP une ville dont [les commerçants _ ] vendent de l’huile (a town of-which the merchants sell oil)

1. a. from VP une ville dont je [parle souvent _ ] (a town of-which I talk often)

         b. from object NP une ville dont j’apprécie [les commerçants _ ].

Corpus studies Using the written corpus French Treebank (Abeillé et al, 2003), we found an overwhelming frequency of extraction out of NP subject (56,5%) over verb extraction (19%) and extraction out of NP object (12,5%) in dont relative clauses. The spoken corpus CFPP2000 (Branca-Rosoff et al. 2012) confirms a high frequency of subject extraction (17,8% of all dont; 63% of all extractions out of NP). These subjects are not internal objects: looking at verbs for DRC with extraction out of subject, 67% of the verbs in the spoken corpus and 47% in the written corpus are transitive (unergative). Furthermore, almost all extractions out of object NP involve a pronominal subject. Linear mixed models also show that dont significantly tends to correspond to a verb complement with a pronominal subject rather than with a nominal subject (z=7,7; p<.001).

Experiment 1 In an acceptability judgement task with 57 participants, we tested extraction out of subject and object NPs with pronominal and nominal subjects for the latter (1b/2b), adding control sentences with coordination (3x2 conditions; 30 items). The participants were asked to rate the sentences on a scale from 1 (bad) to 10 (good).

Extraction out of the subject was judged significantly better than extraction out of the object with a nominal subject NP (1b vs. 1c; p<.05). Extraction out of object NP yielded a significant preference for pronominal over nominal subjects (1b vs. 2b; p<.05). This confirms the results of our corpus studies.

Experiment 2 We ran an Eye-Tracking experiment with the same materials and 31 native French speakers. Total reading times of the verb region confirmed our previous results (controlling for verb length and frequency).

Conclusion The results of the two corpus studies and two experiments show that there is no extra cost of extracting out of a subject NP in DRC. This contradicts structural approaches to subject islands and confirm DLT predictions. The extra cost of a lexical subject when extracting out of an object NP (new discourse referent between dont and the gap) is confirmed as well.

Table 1: Examples of the materials for Experiment 1 and 2