It was generally believed that sentences built up a context for people to predict the upcoming words. With the consensus that people may generate predictions based on the preceding sentence context, researchers tried to manipulate the information load carried by a sentence. Sentence constraint referred to how strongly a sentence frame can lead to a particular final word. It was usually assessed by a normative procedure, in which a group of participants (separated from the main experiment) were requested to complete a sentence frame with the first word appearing in their mind. The percentage of the word used to complete a sentence frame was the cloze probability for that sentence (Taylor, 1953). For instance, assuming that over 90% of the participants filled in the sentence frame “Tom has washed all the dishes on the ______” with the word “table,” the sentence could be defined as a high constraint sentence. In contrast, sentence frames such as “In the forest, there were three _______”
could be filled in with various words (e.g., cabins, pigs, flowers, etc.). If “cabins” was the highest cloze word but was provided by only 35% of the participants, the sentence could be categorized as a low constraint sentence.
How sentence constraint influences the processing of the upcoming word(s) had been extensively discussed.1 In psycholinguistic studies, it was reported that high constraint sentences could facilitate the processing of an expected word. In lexical decision tasks, the response time of an expected word following high constraint sentences was significantly shorter than that of a word following low constraint sentences (Schwanenfluegel & LaCount, 1988; Schwanenflugel & Shoben, 1985).
Similar results could be found in naming tasks, in which participants’ naming time to words appearing in high constraint sentences was faster than those in low constraint ones (McClelland & O’Rgan, 1981). Studies with the eye-tracker technique found a higher skipping rate and less likelihood of regression on a contextually constrained word. If such a word was fixated, the fixation time was shorter than a less constrained word (Ehrlich & Rayner, 1981; Rayner & Well, 1996).
In the field of neurolinguistics, Kutas and Hillyard (1984) was the first ERP study that manipulated the levels of sentence constraint to investigate the effect of word expectancy during sentence processing. In this experiment, sentence frames of various constraining levels (high, medium, low) were terminated by words of different cloze probabilities (high, medium, low). The results showed that high constraint sentences with high cloze words elicited a board positive wave whereas other conditions elicited negative-going waves (N400)2. The brainwave differences indicated that participants responded differently when the presented word matched their prediction. It should be noted that sentence constraint and cloze probability
1 A considerable amount of researches have manipulated the sentence context to bias one of an ambiguous word’s meanings (Rayner, Cook, Juhasz, & Frazier, 2006;
Pickering & Frisson, 2001). However, this is beyond the scope of the current study and will not be discussed here.
2 N400, a component peaking around 400ms post-stimulus, had been shown to vary systematically with processing semantic information. It was actually more sensitive to cloze probability rather than the degree of sentence constraint. We will discuss the effect of cloze probability in Session 2.2.
might not be mutually independent. In fact, only highly constrained sentences had the possibility of being terminated by a very high cloze word.
To tease the factors of sentence constraint and cloze probability apart, Federmeier, Wlotko, Ochoa-Dewald, and Kutas (2007) adopted a fairly smart design with high constraint and low constraint sentences. In the experiment, each sentence was either terminated with its highest cloze word (the best completion) or with an unexpected but semantically plausible word of zero cloze probability. The reasoning was that, although the cloze probability of the best completions could not be easily controlled between two constraint levels, the cloze probability of the terminal word could be matched (i.e. zero) in the unexpected but semantically plausible conditions.
Any differences between the unexpected words could be attributed to the effects of sentence constraint per se. The ERP recording revealed that only high constraint sentences with an unexpected completion elicited a positive brainwave at the frontal electrode site. Such an effect, as termed frontal P600 by the authors, was found neither in the best completion in high constraint sentences nor in both completion types in low constraint sentences. The frontal P600 was different from a canonical P600/LPC, which was usually largest at central-parietal electrode site and considered to be an indication of syntactic reanalysis (Osterhout & Holcomb, 1992). Federmeier et al. (2007) interpreted the frontal P600 positivity as “an appreciation to mismatches and/or the allocation of resources necessary to revise a prediction” (p. 8).
Since Federmeier et al. (2007) had demonstrated sentence context could affect the processing of the upcoming word, to further investigate to what degree can people predict an upcoming item, Thornhill and Van Petten (2012) manipulated the semantic relatedness of the sentence completions. In the experiment, high and low constraint sentences were completed by the best completion (BC), by an unexpected
near-synonym of BC (related), and by an unexpected word semantically unrelated to the BC (unrelated). The cloze probability of related condition and unrelated condition were carefully matched. Examples of high and low constraint sentences respectively were “He was afraid that doing drugs would damage his brain (BC)/mind (related)/reputation (unrelated)” and “Penelope started to assemble her new bicycle but was missing the wheels (BC)/tires (related)/instructions (unrelated).” The results revealed that N400 was not only sensitive to target word expectedness (as indicated by cloze probability) but also semantic relatedness to the BC. In contrast, the late frontal positivity, or the frontal P600 in Federmeier et al. (2007), seemed to be more sensitive to lexical form expectations, as it was reduced only in high constraint BC condition. Other conditions did not have significant differences in terms of the amplitude of the late frontal positivity. Thornhill and Van Petten (2012) suggested that readers could predict both conceptual and lexical information of the upcoming words, although the processing of conceptual and lexical mismatches could be subserved by distinctive neural networks.
Both Thornhill and Van Petten (2012) and Federmeier et al. (2007) observed a late frontal positivity in high constraint sentences with unexpected sentence completions. In addition, they both agreed that the late frontal positivity was an indication of disconfirmed prediction. Thornhill and Van Petten (2012) took a step further, by attributing the late frontal positivity to unexpected lexical forms rather than unexpected concepts. However, a major difference arrived when comparing the low constraint conditions in the two studies. In Thornhill and Van Petten (2012), low constraint sentences elicited the late frontal positivity, but such positivity was absent in Federmeier et al. (2007). As for now, it seems that little consensus has been reached on the predictability effect of low constraint sentences. Still, researchers
generally agreed that high constraint sentences provided the context for people to predict the upcoming word. If the presented word did not match the prediction, a late frontal positivity would be elicited. In the current study, we would use high constraint sentences as the experiment materials. We will discuss more about the late frontal positivity in Session 2.2.