Decomposition makes things worse: A discrimination learning approach to the time course of understanding compounds in reading by (Eberhard Karls University, Tübingen, Allemagne)
The current literature on morphological processing is dominated by the view that reading a complex word is a two-staged process, with an early blind morphemic decomposition process followed by a late process of semantic recombination (Taft, 2004; Rastle and Davis, 2008a). Various behavioral and magneto- and electroencephalography studies suggest semantic recombination would take place approximately 300-500 ms post onset of the visual stimulus (Lavric et al., 2007). However, eye-tracking studies show that both simple and complex words are read at a rate of 4 to 5 words/second (Rayner, 1998). We report an eye-tracking experiment tracing the reading of English compounds in simple sentences. For about 33% of the trials, a single fixation sufficed for understanding the meaning of the compound. For such trials, the meaning of the compound was available already some 140 ms after the eye first landed on the modifier. All first fixations also revealed an effect of the semantic relatedness of the modifier and head constituents, gauged with a latent
semantic analysis (LSA) similarity measure. These results indicate a much earlier involvement of semantics than predicted by the first-form-then-meaning scenario. Second and subsequent fixation durations revealed that at later processing stages very different semantic processes were involved, gauged by modifier-compound and head-compound LSA similarity measures. Computational modeling of the first fixation with naive discrimination learning (Baayen et al., 2011) indicated that the early (and only the early) semantic effect arises due to the model's connection weights' sensitivity to the collocational co-occurence statistics of orthographic and semantic information carried by word trigrams. We understand the LSA effects arising at later fixations as
reflecting semantic processes seeking to resolve the uncertainty about the targeted meaning that arises as an unintended and time-costly side effect of later fixations causing the head's meaning to be co-activated along with the compound's meaning. Instead of viewing blind morphological decomposition as the gateway through which meaning can be reached, we think that when the meaning of the head becomes available, due to the (non-morphological) nature of visual information uptake when the initial landing position of the eye is non-optimal, understanding comes with greater cognitive costs: Decomposition makes things worse. We speculate that the late semantic effects in the electrophysiological literature, especially those around the N400 time window, reflect late semantic cleaning operations.