To illustrate the properties of these two control mechanisms in more detail, we will now first set out what exactly triggers a regressive eye movement within the Information Gathering Framework. Note that this procedure clearly differs from the one proposed by the FC model because it tries to also model the time course of language processing and the limited focus of attention. In particular, the FC model only knows one scenario that causes a regression, namely, if evidence from input later in the sentence causes low con- fidence of precious entities, whereas the Information Gathering Framework proposes two different regression scenarios.
53 First of all, recall that the focus of attention is restricted to word n and word n+1, which we will refer to as word 5 (W5) and word 6 (W6) respectively, according to Figure
4. Now, let us consider what happens when the eyes fixate on W6.
First, the letter string of W6 is perceived and the lexical quality level is computed.
Because this takes a certain amount of time, the quality level (the amount of lexical in- formation) of W6 is typically increasing during the fixation on W6. In addition, the confi-
dence level is computed by matching the features of the lexical representation with the predictions according to the production rules. Thus, the confidence level is also typically increasing during a fixation. If it reaches the forward threshold, an eye movement to the next word (W7) is performed.
Second, at the same time the quality level of W5, that peaked right after the eyes
moved to W6, is now slowly decreasing due to interference and decay. The computation
of the confidence level on the basis of this lexical representation, by contrast, stops and the confidence level remains stable. Note that the confidence of W5 already reached the
forward threshold because this triggered an eye movement to W6. But the lexical repre-
sentation is now used to match also higher-order predictions that need more time to be computed. Thus, the confidence level of W5 is typically also increasing during a fixation
on W6 and reaches the backward threshold. If this backward threshold is reached before
the confidence level of W6 reaches the forward threshold, everything is fine and an eye
movement to W7 is performed (see Figure 5, Pattern 4).
There are two scenarios, however, where this forward walk through the sentence is disrupted. First, if the computation of the confidence level of W5 reveals great difficul-
ties that question the confidence into a word’s identity so that the confidence level of W5
even falls under the forward threshold. This causes a regression, independently of the confidence level of W6.18 We assume that this happens especially due to higher order
parsing problems that do not allow for an integration of the current word into the adopted sentence structure.
Second, if the computation of the confidence level of W5 takes more time because
the predicted evidence is missing or more information has to be taken into account, the
18 Note that also the computation of the confidence level of W6 may detect very early in-
tegration difficulties (before the forward threshold is reached). In this case, either fixation dura- tion on W6 increases or the information is used for the computation of the confidence level of W5
because both words are in the focus of attention. In this case, even information about W6 may
cause that the confidence level of W5 falls under the forward threshold and a regression is trig-
54 confidence level of W6 may reach the forward threshold before the confidence level of
W5 reaches the backward threshold. This also triggers a regression. We assume that this
happens especially at the end of a sentence where the sentence is evaluated as a whole (see chapter 3.1.3).
To illustrate the two scenarios, consider, for example, one of the stimulus sen- tences again used in the famous experiment by Frazier & Rayner (1982):
(3) Since Jay always jogs a mile and a half seems like a very short distance to him. When the eyes fixate on the word “seems”, the visual string is received, the lexical infor- mation is retrieved from the memory and the lexical quality level increases. Because the string is acceptable on lower linguistic levels (since it is a well-formed and highly fre- quent word of English, for example), the confidence level reaches the forward threshold soon and a progressive saccade to the next word “like” is performed. However, during the fixation of “like”, the computation of the confidence levels of both words (like and seems) is performed in parallel. For the word “seems” the production rules now yield an error on the syntax level because the word cannot be integrated into the current sen- tence structure. As a response, the confidence level for the word “seems” decreases and falls under the forward threshold which causes a regressive eye movement, inde- pendently of the confidence level for the word “like” (scenario 1).
For the second scenario, consider a fixation on the last word of the sentence: “him”. Here the confidence levels of “him” and “to” are computed in parallel and no vio- lation of production rules is detected. However, because the eyes parafoveally receive the punctuation which signals the end of the sentence, the reader evaluates the meaning of the whole sentence in order to increase the confidence level of the word “to”. This may lead to a delay because the reader might have expected more evidence that was not given in the course of the sentence (for example, more information about Jay). In the mean- time, the confidence level for the word “him” has reached the forward threshold, alt- hough the confidence level for “to” has not reached the backward threshold. This also causes a regressive eye movement in response (scenario 2).
In both cases, however, a regressive eye movement is performed because the available information about current word identities is assessed to be insufficient for the current sentence interpretation. Thus, a regression always aims to gather additional ev- idence about a word’s identity and to increase the level of lexical quality of previous
55 words by making a shift of attention. This in turn normally also increases the confidence level of these words.
Figure 5: Potential patterns of confidence levels. Each pattern represents the confidence levels of six words (W1 to W6) during a fixation on W6. The blue line shows the backward threshold and the orange line the
forward threshold.