Recovery of elliptical structures: electrophysiological evidence for retrieval and integration processes

Elliptical sentences (e.g. “John bought a book in the shop, and Sally too.”) are recovered by a process of retrieving and integrating earlier mentioned information (here: “bought a book”). Based on behavioural results it has been proposed that ellipsis may be resolved by inserting a copy of the missing structure (e.g. “bought a book”), predicting a low processing cost – regardless of the size of the antecedent [1]. This mechanism may however not be applicable for all ellipsis types (e.g. ‘gapping’) [1,2]. While it has recently been proposed that ellipsis may be constrained by structural and discourse conditions [3], it is unclear how a copy mechanism would accommodate non-syntactic conditions.

Some authors have argued that, in general, accessing and copying more structure should take more time, proposing an alternative view of ellipsis resolution that involves step-by-step structure building, while a more fully interpreted discourse representation of the antecedent is accessed and integrated. Implemented as a memory based content-addressable pointer mechanism it would involve direct access of the antecedent, suggesting that the speed of interpreting the ellipsis does not depend on antecedent complexity [4]. Consequently, the proposed mechanisms predict comparable behavioural results with respect to structural complexity effects which might also be due to issues relating to experimental design [5]. In the present study, we explored the possibility to differentiate between the two mechanisms by means of electrophysiological data.

Few event-related brain potentials (ERP) experiments on verb-gapping have reported that missing structure is detected quickly, while a P600 component has been found and interpreted as a reflection of an integration process of a ‘missing’ verb and remnant object [6,7]. In the current ERP study we investigated the effects of the structural size of the antecedent on retrieval and integration processes. In accordance with a copy mechanism, we hypothesised that the onset of ERP signatures relating to accessing and copying missing structure may vary as a function of structural size; upon retrieval a fully fledged structure may facilitate the integration process. A pointer mechanism could account for the reverse. Early ERP signatures of retrieval may be fixed since the antecedent is directly accessible. On the other hand, integration processes may operate on representations of various types, as discourse information is integrated in incrementally built up structure. The latter will predict ERP variability relatively late in the time-course.

Using a word-by-word reading task, we studied the online processing of a sub-type of gapping, i.e. ‘stripping’, in Dutch. We recorded ERPs while 30 Dutch participants read sentences containing stripping constructions (equivalent to “and Sally too”) interspersed with filler sentences. Critical measure point was the onset of the word ‘ook’ (too). The conditions differed with respect to the size of the elided structure: verb-noun (condition A), verb-adjective-noun (condition B), adjunct-verb-noun (condition C) or adjunct- verb-adjective-noun (condition D). Every stimulus sentence was followed by a comprehension question. We counterbalanced which sentence parts (including elliptical phrases) were questioned. To control for individual differences, participants also performed a working memory task.

Since we were interested in onset and duration differences, we conducted repeated measures ANOVAs for 100 ms time windows. We observed a significant effect of the factor condition on the mean amplitude between 300 and 400 ms after onset of ‘ook’. A Bonferroni post-hoc test revealed that condition A did not differ from B. Compared to condition A, a significant centro-frontal positive deflection could be established in both conditions C and D. Between conditions C and D, no significant difference was apparent. However, after 400 ms the positivity appeared to be sustained and broadly distributed in condition D. The participants’ performance on the comprehension questions did not differ between conditions. We found a weak, non-significant positive correlation between the accuracy on the comprehension questions and the scores on the working memory test.