According to the FC model, the reader computes a confidence level of a particular word on the basis of its language model. This confidence level of a word is updated as more information in the course of the sentence is given. In particular, the FC model proposes
language experience) but that he has a very detailed representation of words that are in specific fields of which he is an expert in.
46 that additional information may also cause the confidence in a previous word’s identity to fall. In response, a regressive saccade to this particular word is triggered, whenever its confidence level declines under a certain threshold.
As a consequence from the underlying bigram frequency model and the focus on reducing noisy visual input, the FC model does not view the computation of confidence levels as a matter of language processing. In particular, the FC model does not specify the lexical representation of a word in more detail, so that linguistic computations do not affect fixation durations.
This is unquestionably a very unrealistic assumption because reading is not only a visual decoding task but rather aims at the comprehension of a linguistic message. As already discussed above, the major determinant for fixation durations seems therefore to be linguistic processing instead of visual perception. In particular, because language processing is, like all cognitive processes, a time consuming task, decades of neuro- and psycholinguistic research have focused on determining the exact time course of language processing. This time course of language processing is, however, likely to correspond (at least partly) with the language hierarchy, with lower linguistic levels being processed earlier than higher linguistic levels.
Within the Information Gathering Framework it is therefore assumed that the computation of confidence levels (as with the computation of the lexical quality levels) is based on linguistic processing and takes a certain amount of time. During this time, the confidence level of a word typically increases (asymptotically approaching but never reaching the full confidence level), because more supporting evidence is given from the information of the lexical representation (see Figure 3 for a schematic illustration). Ac- cording to the limited focus of attention which we will discuss in more detail in the next section, the computation of the confidence level of word n is restricted to the time the eyes fixate on word n and word n+1. After the eyes have moved to word n+2, no further computation of word n’s confidence level can take place and the confidence level remains stable over time.
Importantly, the Information Gathering Framework follows the “one-system hy- pothesis” of grammatical theories and language processing models (see Lewis & Phillips, 2015, for a recent discussion) claiming that the time course that is proposed within the framework corresponds to the time course of linguistic processing and in turn to the lin- guistic hierarchy. Thus, the earlier in time, the lower the linguistic level of the lexical representation, of which the computation of the confidence level is based on.
47 A challenging question, however, is how the confidence level is exactly computed. In the FC model, the confidence level is calculated on the basis of the noisy visual input by using a Bayesian inference term that computes the distribution over possible identi- ties. Importantly, the FC model uses only bigram frequencies and no further linguistic information. Whereas this architecture of the FC model was chosen to allow for a formal- ized integration into a computational account, the situation within the Information Gath- ering Framework is much more complex because it explicitly claims that language pro- cessing is the major factor of eye movement control in reading. However, to map the complexity of language processing in full detail is beyond the scope of the present frame- work. Thus, the mechanisms proposed here are a simplification, but may be elaborated in more detail by future research.
For the current purpose, it is assumed that the confidence level is computed by matching the features of the lexical representation with the predictions of former sen- tence material on the basis of explicit production rules (Newell, 1973). These production rules represent all procedural knowledge (grammatical knowledge) and set condition– action pairs. For example, if an inanimate noun (e.g. the table) is encountered as the ini- tial argument in an English sentence (condition), the production rules predict that a verb (action) will follow in the course of the sentence. More precisely, they predict that this verb should agree with the argument in number (singular), comes with an inanimate subject, and so on. If a verb like talks is encountered next, this leads to a violation of production rules because talks requires an animate subject. On the other hand, if a pro- noun like the word which is following, it induces a relative clause. In this case, the pro- duction rules are not violated and the action (the expected verb) is simply postponed. Also, not every condition-action pair is mandatory; some pairs are just optional (e.g., the indirect object of verbs like write: He writes a letter (to his father)). If the evidence pro- vided by the lexical representation matches the predictions made on the basis of the pro- duction rules, a high confidence level is computed. If the production rules are violated by contrast, it leads to a low confidence level. Accordingly, if the context is highly predictive, less lexical information and thus less time is needed to reach a certain level of confidence which results in shorter fixation durations.
Crucially, there are two different scenarios that trigger a regressive eye move- ment. In the first scenario, the computation reveals that the confidence level of word n falls under the so called forward threshold (that defines the level of confidence that is needed to proceed), although the eyes have already moved to word n+1. This happens
48 if the predictions of the production rules are violated by the information from the lexical representation. In the second scenario, the computation causes the confidence level of word n to not reach the so called backward threshold (that prevents a regression from happening) within the fixation of word n+1. This happens if the optional predictions of the production rules are not met or other expected supporting evidence for the identity of a word is not given. We will explain these two scenarios in more detail in section 3.1.7, after we have outlined the assumptions about the limited focus of attention and the two control mechanisms.