GONZALEZ NUEZ ALBA

The FC model proposes that because word identification is based on noisy visual infor- mation,“word recognition may be best thought of as a process that is never ‚completed’” (Bicknell & Levy, 2010, p. 1170). Accordingly, the reader uses the noisy visual input to compute the probability of the identity of a word on the basis of its language model (that is organized by frequency), but always a level of uncertainty remains.

Although we agree on the assumption of incomplete word recognition, we may ask whether noisy visual information is in fact the major determinant of word identifica- tion that may explain the variability in fixation durations. In particular, there exists con- vincing evidence that the decoding of visual information occurs very rapidly. That is, even when a word is presented for only 60 ms and masked afterwards, reading behavior is unaffected (Ishida & Ikeda, 1989). It seems therefore more plausible to assume that word identification is mainly affected by the retrieval of the lexical information (as also proposed by the SWIFT and E-Z Reader model) instead of problems with the visual per- ception. Thus, longer fixation durations would (primarily) correlate with an increase of lexical information and not with an increase of visual information, because the visual information is decoded rapidly.

This assumption, of course, requires a more complex language model than the simple bigram frequency model underlying the FC model that has limitations in the face

11 _{Here we propose an account that is word-based as most models of eye movement con-}

trol are (SWIFT, E-Z Reader etc.). But we think that this a simplification though because often single words can be better viewed in combination with other words, either as a fixed unit (as in chunks) or as a syntactical phrase that has a common function within a sentence (e.g. in noun phrases). This view is often held in sentence processing accounts that focus on phrases rather than words.

43 of the full complexity of linguistic processing during sentence reading as well (as dis- cussed in chapter 2.2.3). Specifically, the lexical representation stored in the memory has to be viewed as an (theoretically) infinite bundle of features, containing information about the word’s orthography, phonology, meaning, morpho-syntax as well as its con- stituent binding preferences (see Figure 2 for a schematic illustration; c.f. also Perfetti [2007] who introduced this idea as the concept of lexical quality in order to explain dif- ferences in language skill between individuals). Thus, whereas the FC model proposes that increased fixation times enable the reader to decrease noisy visual input, we make the claim within the Information Gathering Framework that increased fixation times pri- marily enable the reader to retrieve more information about a word’s identity from the lexicon.

Figure 2: A schematic illustration of the lexical representation of a word that is stored in the memory. This representation comprises the five linguistic levels orthography, phonology, meaning, morpho-syntax and constituent binding (see also Perfetti, 2007). Each of these levels consists of several (or infinite) more fea- tures. The lexical quality level describes the amount of information about a word that is currently retrieved from the lexicon.

Whereas the term ‘lexical representation’ refers to the stored entry of a word in the lexicon, the term ‘lexical quality level’ describes the amount of information about a word that is currently retrieved from the lexicon. Typically, the amount of information (and thus the lexical quality level) continuously increases during a fixation, because a fixation allows for the retrieval of lexical information on the basis of the visual input12_.

12 _{This does not mean that the whole time during a fixation is needed to encode the visual}

input (because this happens very fast, as mentioned earlier) but that the visual input allows for the retrieval of the lexical information.

44 However, once the eyes have moved to the next word, no additional information can be received13 _{and the quality level is then continuously decreasing over time due to inter-}

ference from other words and due to a decay14 _{(Lewis & Vasishth, 2005; see Figure 3 for}

a schematic illustration). Also note that the lexical quality level of a word (as the confi- dence level) is never reaching the full quality level because the retrieval of the infor- mation from the lexical entry can by definition never be completed.

Figure 3: Schematic illustration of the confidence / quality level of a single word during a typical sequence of two progressive saccades: Whereas the confidence level is continuously increasing and asymptotically approaching the full confidence level, the quality level decreases due to interference and decay after the eyes moved to the next word. Legend: purple = quality level, green = confidence level, orange = forward thresh- old, blue = backward threshold, S1 = saccade to word n+1, S2 = saccade to word n+2, t = time.

It is obvious, however, that the lexical quality level depends on the lexical repre- sentation that is stored in the memory. Thus, the size and the specification of this lexi- con15 _{in particular, may affect the process of word identification in two ways. First, the}

13 _{This assumption might be problematic because there is evidence showing that the}

masking of the immediately preceding word does also affect reading behavior (e.g., Rayner, Well, & Pollatsek, 1980). This suggests that the visual information of word n-1 is still used for word identification. However, it is unclear if this information affects the retrieval of lexical representa- tion (as indicated by the lexical quality level) or processes that primarily require attention (as the computation of confidence levels) which are assumed to continue after the eyes have moved to the next word.

14 _{It could also be that an eye movement within a word is performed (intra saccade). In}

this case, the quality level of a word is also increasing after the have eyes moved. But since we are focusing on inter-word saccades because intra-word saccades have to be viewed as functionally different from inter-word saccades (see chapter 1.1), we just aggregate all saccades within a word to a single inter-word saccade (thus talking in fact about first pass times, as most models of eye movement control in reading do).

15 _{The size and specification of the lexicon do highly interact but are not the same. Thus,}

45 lexicon affects the amount and accuracy of information about a word’s identity that can be retrieved from the visual input. Thus, a word representation that is stored in the lex- icon may consist of a few features only if the word is not well known by the reader. Al- ternatively, it may consist of many features if the reader has a very detailed representa- tion of a word (consider, e.g., the representation of technical terms). Also, the associated features could be correct or not (e.g., in phonology). Second, the lexicon affects the time it takes to retrieve this information from the lexicon because the word entries in the lex- icon are sensitive to frequency (Forster & Chambers, 1973; Marslen-Wilson, 1990). That is, the more frequently an entry or its features are retrieved, the more easily (and swiftly) this information can be retrieved again.

In addition, the size and specification of the lexicon is also assumed to vary among individuals, depending on individual language exposure, language skill and the language under consideration. Thus, given a certain amount of time, the lexical quality level of a word is expected to vary amongst readers, as well as the time taken to reach a certain level of quality.

Note also that the lexical quality level is highly correlated to the level of confi- dence, but these two parameters are not the same. A poor reader could have a high con- fidence in a word’s identity although it is ambiguous (e.g., in meaning). But due to a small lexicon which implies a representation of a few features only, the reader is not aware of these alternative interpretations. Accordingly, a proficient reader could have a low con- fidence into the same word’s identity because he takes into account several potential ambiguities that the poor reader is not aware of. In addition, a highly predictive context may also effect that less information (and thus a lower lexical quality level) is needed to confirm this prediction and reach a certain level of confidence. This explains why fixa- tions on highly predictive words are shorter than those on unpredictable words (see next section).