Does the parser exclusively use structure-sensitive search in reflexives? Evidence from Chinese reflexives

Much recent work (Xiang et al 2009, Phillips et al 2009, Dillon et al. 2010) has argued, inter alia, that the human parsing mechanism employs an intelligent search process that uses grammar-based constraints. For example, in (1) the reflexive herself resolves to the c-commanding antecedent `woman' (`grammatically accessible' antecedent) without experiencing any confusion due to the NP Katie, which matches herself in gender marking but does not c-command it. As Phillips et al (2009) put it: "we tentatively suggest that argument reflexives are immune to interference from structurally inaccessible antecedents because antecedents are retrieved using only structural cues".

Although the parser may use an intelligent search process using structural cues, the retrieval process is also subject to interference arising from cue-based retrieval difficulty (Lewis & Vasishth 2005). I.e., a grammatically inaccessible noun like Katie in (1) would cause increased difficulty in resolving the antecedent of the reflexive, due to the gender match between the reflexive and the inaccessible noun. 

Support for the Phillips et al structure-sensitive search view is necessarily based on null results, usually with small sample sizes (low power).  An obvious question is: would interference effects emerge if power were high enough? Using a larger sample size (n=120), we investigated the opposing claims of the structure-sensitive-search versus the cue-based retrieval account. 

In Chinese structures like 2, the antecedent 'the opposition leader' is the only legal antecedent for the reflexive ziji, `self' (which requires an animate antecedent). In Chinese the antecedent isn't necessarily clause-bound: 2a,b.  Under the structure-sensitive search view, the parser should never consider an intervening noun like kangyizhe, `protester', as an antecedent because it is inside an adverbial phrase and cannot c-command the antecedent ziji; this predicts no reading time difference between cases where the intervening noun is `protest' (2a,c) versus `protestor' (2b,d). By contrast, the cue-based retrieval account predicts slower reading time in (2b,d) vs (2a,c). 

We found an interference effect (Tables 1, 2) as predicted by the cue-based retrieval account, and inconsistent with the exclusively-structure-sensitive search account: there was a main effect of interference at ziji and the word following it that was driven by the non-local antecedent conditions 2a,b. These results suggest that, although the parser may well consult grammatical constraints when resolving dependencies, i.e., the parser may rely on structure-sensitive search, cue-based retrieval interference is unavoidable due to a partial feature match (here, animacy match) between the intervening noun and reflexive. 

1. The woman that {Fred|Katie} treated in the military hospital introduced herself to all the nurses.

2. a. Long-distance, no interference:
      fanduipai-lingxiu_i ... [zhege-shengming [zai kangyi shikong de-shihou] ... ziji_i ...]
      opposition-leader ...   [this-announcement [at protest  out of control time]...ziji...]
     `The opposition leader indicated that this announcement warned his party members when the protest/protestor was out of control.'
   b. Long-distance, high interference:
      fanduipai-lingxiu_i ... [zhege-shengming [zai *kanghizhe*  shikong de-shihou] ... ziji_i ...]
      opposition-leader ...   [this-announcement [at *protestor* out of control time]...ziji...]
   c. Short-distance, no interference:
	  zhege-shengming  ...  [fanduipai-lingxiu_i [zai kangyi shikong de-shihou] ... ziji_i ...]
	  this-announcement ... [opposition-leader [at protest  out of control time]...ziji...]
   d. Short-distance, high interference:
	  zhege-shengming ...   [fanduipai-lingxiu_i [zai *kangyizhe* shikong de-shihou] ... ziji_i ...]
	  this-announcement ... [opposition-leader [at *protestor* out of control time]...ziji...]


Table 1: Reading times at critical and post-critical region. 
           a     b     c     d               Differences:
ziji     410.1 447.9 410.5 414.8             b-a: 37.8 ms  d-c: 4.3 ms  
ziji+1   371.1 387.6 362.5 371.9             b-a: 16.5 ms  d-c: 9.4 ms

Table 2: Linear mixed model with participants and items as crossed random factors.
       contrast    coef       se       t            contrast   coef       se        t
ziji   Locality   -0.0258     0.0134  -1.9*  ziji+1 Locality  -0.02367    0.01046  -2.3*
	   Int         0.0273     0.0134   2.0*	        Int        0.02349    0.01046   2.2*
	   Loc x Int  -0.0265     0.0134   2.0*	        Loc x Int -0.00987    0.01046  -0.9 

Refs:
Dillon et al. submitted, 2010.
Lewis & Vasishth. 2005 Cognitive Science.
Phillips et al 2009. Language and Linguistic Compass.
Xiang et al 2009. Brain and Language.





