von der Malsburg, T. & Vasishth, S.
Deparment of Linguistics. University of Potsdam. Potsdam, Germany.
Over the last decades there have been several proposals about how the language system repairs an analysis of a sentence when material is encountered that is incompatible with the interpretation of the sentence maintained so far. In reading, this repair process is often accompanied by multi-fixation backwards movements of the eyes (regression scanpaths). The evidence from two studies that analyzed these scanpaths is inconclusive: Meseguer, Carreiras, Clifton (Mem&Cog, 2002) found support for an intelligent and targeted repair mechanism (c.f. Frazier, Rayner, CogPsych, 1982), but Mitchell, Shen, Green, Hodgson (JML, 2008) also found evidence for an influence of low-level visual factors on regressive eye movements, indicating that the parser is not in full control of eye movements during syntactic reanalysis. Meseguer et al. used Spanish sentences with a temporary attachment ambiguity where an adverbial phrase could attach high or low. We ran a Spanish follow-up study (n=70) that extended their design to include (among other things) an unambiguous baseline condition. Instead of analyzing local measures like fixation durations and transition probabilities we used a new method that directly quantifies spatio-temporal differences in the full fixation patterns (von der Malsburg, Vasishth, CUNY, 2010). Our results show that, contrary to what was believed previously, there is not just one reanalysis mechanism. Instead, readers use one of three different strategies when they encounter the disambiguating material: (i) diagnosing the problem by rechecking the disambiguating word, presumably followed by covert reanalysis, (ii) searching for cues, reflected by erratic fixation sequences, (iii) rereading of the sentence. Readers differ markedly in how they orchestrate these strategies. We hypothesize that each strategy corresponds to a different degree of confusion of the parsing system. In general, these results show that the scanpath approach is capable of uncovering processing strategies where traditional methods have failed to do so.