French object relatives: evidence against DLT but not entirely explained by frequency

International audienceFrench object relatives: evidence against DLT but not entirely explained by frequency Céline Pozniak, Barbara Hemforth &amp; Anne Abeillé (LLF, CNRS, Univ. Paris-Diderot, Labex EFL) celine.pozniak@gmail.com The literature in relative clause processing mainly presents differences between subject and object relative clauses (SRs and ORs). In our study, we will also compare two constructions for ORs: one with preverbal subject (ORsnoinv) and one with postverbal subject (ORsinv). Inversion is only possible with nominal subjects, not with pronominal subjects. (1) OR with preverbal subject<br /> Le médecin que l&#39;avocat voit ___. The doctor thatobj the lawyersubj sees.<br /> Example (1) shows that the linear distance is shorter in ORsinv , so it should be easier to process than ORsnoinv according to the Dependency Locality Theory, now DLT (Gibson, 2000).<br /> Study 1 We designed an experiment to test linear distance-based theories in French in contexts with high anticipation of restrictive relative clauses.<br /> Experiment We ran a Visual World Eye-Tracking experiment (5 items per condition) with 32 native French speakers. We tested reversible SRs, ORsnoinv, ORsinv (1-3). The participants listened to a sentence while viewing 2 pictures with the same 3 characters each performing different actions. They had to find the correct picture according to the sentence. One picture was only compatible with an SR interpretation, the other one only with an OR interpretation. Results Mixed linear models showed a significant SR advantage over the two ORs (ps&lt;.01). Participants looked at the right picture in later time windows for ORinv than for ORnoninv (ps&lt;.01), against linear distance accounts. These results seem, however, to go against the intuition that ORinv are relatively frequent in French, but possibly only under specific conditions. Study 2 We therefore ran a corpus study using the French Treebank (Abeillé et al, 2003) to analyze factors for facilitating inverted and non-inverted RCs. Corpus study We analyzed a corpus of the two ORs by looking at the semantics of the verb (+/- agentive), the length of the subject and the verb in number of syllables, and the number of arguments in the relative. We analyzed the role of these factors for the choice of the order in the relative using logistic regression models. Results Logistic regressions show that the two ORs don&rsquo;t differ significantly in frequency, excluding a simple frequency-based explanation (94 ORsnoinv and 90 ORsinv). However, ORsinv are actually preferred over ORsnoinv when the subject is longer, and the verb is shorter and has a non agentive meaning (ps&lt; .01). Conclusion In our Eye-Tracking experiment, ORsinv were harder to process than ORsnoinv, contradicting the DLT. However, the corpus study showed that general frequency cannot be the factor explaining this contrast. The verb semantics and the length of the subject and verb play a role in the choice of relative. This means that ORsnoinv and ORsinv are not just alternatives but used in specific contexts. The context in the experiment disfavored the use of ORsinv because of the use of agentive verbs. In the right context, ORsinv should even be preferred and acceptability judgment studies on this matter are currently on the way. Our results are highly compatible with semantic/pragmatic accounts in relative clause processing (Mak et al, 2006; Traxler et al, 2002).&nbsp