Chapter 2 Literature Review
2.2 Parafoveal processing
2.2.4 Reading models and semantic preview benefit
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
definition of preview time, increased the preview benefits. In another study by Tsai, Kliegl, and Yan (2012), where the operational definition of preview time was pretarget GD, preview benefits also increased with longer preview time.
Details of some of the studies mentioned above will be discussed later. For now, it seems clear that preview effects can be modulated by the factors of preview time and preview space. Further, according to how these factors influence the preview effects – by increasing the reading times with unrelated preview or by decreasing the reading times with related or identical preview, one could conclude whether such preview effects are due to preview benefit from related features or preview cost from unrelated features. Before further examining the studies in semantic preview benefit, the theoretical importance of such effect regarding the reading models will be discussed in the next section.
2.2.4 Reading models and semantic preview benefit
With accumulation of eye movement data, models have been proposed to fit and explain the data from our reading behavior. The two most successful and popular models are E-Z Reader (Rayner, Li, & Pollatsek, 2007; Reichle, Liversedge, Pollatsek,
& Rayner, 2009; Reichle, Pollatsek, Fisher, & Rayner, 1998; Reichle, Pollatsek, &
Rayner, 2006, 2007; Reichle, Warren, & McConnell, 2009) and SWIFT (Engbert,
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Longtin, & Kliegl, 2002; Engbert, Nuthmann, Richter, & Kliegl, 2005), which belong respectively to serial attention shift (SAS) models and attentional gradient (GAG) models. According to E-Z Reader, attention is allocated at each word during reading, while in SWIFT, attention is distributed around the fixation point, where the level of activation of each word being attended rises simultaneously, since they are processed in parallel.
Although detailed specifications of each type of models are not the issue here, descriptions of their architectures are discussed for further discussion in semantic preview benefit. In E-Z Reader, detailed mechanism has been proposed about the time course and sequence of stages of visual processing, lexical processing, of attention shift, and of saccade planning. In E-Z Reader, lexical processing is assumed to be word-wise serial. A word is being processed when attention, not fixation point, is allocated to that word. There are two stages of lexical processing, L1 and L2. L1 is the early stage of lexical processing and is generally associated with processing of phonological and orthographic information. The completion of L1 will initiate L2, where deeper processing such as semantic processing takes place. The speed of each stage is influenced by factors such as word frequency, contextual information, and deviation of the word from the fixation point. While lexical processing is strictly serial, eye movement control in E-Z Reader is modulated by the state of lexical
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
processing. E-Z Reader assumes two cascading stages of saccade planning as well.
The labile stage M1 can be cancelled and replaced by new saccade planning, but when this saccade planning enters the non-labile stage M2, the planned saccade will have to be executed at the moment when M2 is done, before any new saccade planning can take place. A new saccade planning is triggered by completion of L1 of a certain word n, termed L1n. This new saccade planning aims at the next word n+1, termed M1n+1. Following the architecture of E-Z Reader above, when the eyes fixate on a certain word n but the first stage of lexical processing on the next word n+1 (L1n+1), for example, completes before first stage of saccade planning to the next word (M1n+1) is done, M1n+1 will then be canceled and replaced by M1n+2, to the even next word. However, when L1n+1 completes at the time when the saccade planning to the next word has entered the second stage (M2n+1), it cannot canceled this saccade until M2n+1 is done and the saccade is carried out. According to E-Z Reader, the first situation is the case of skipping (of word n+1), and in the second situation, preview benefit occurs. In the second situation, lexical processing of word n+1 completed before the eyes fixate, but not skip, the word, and that shall reduce the time spent on word n+1.
SWIFT model, on the other hand, assumes distributed attention, and the words within the attention window are being processed at different rate. The speed of word
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
processing, which results in the rises of what SWIFT terms activation levels, is decided by various factors, including the level of attention distributed to it. What differs SWIFT most from E-Z Reader is the temporal and spatial decision regarding saccade planning. In E-Z Reader, saccade planning has a specific goal at the beginning of M1, which is triggered by level of lexical processing (completion of L1).
In SWIFT, however, saccades are autonomous generated without specific targets. Two stages of saccade programming are also suggested by SWIFT, first one labile followed by the non-labile stage. If there is no new saccade programming that intervenes and cancels the current labile saccade programming, at the end of the labile stage, saccade target will then be decided based on the activation level of each word.
As for temporal variation in SWIFT, the base for saccade latency is the stochastic process in saccade generation, modulated by the intended saccade amplitude. Since saccade target is decided with the completion of labile stage, this modulation influences only the length of non-labile stage. Furthermore, saccade latency is modulated by foveal inhibition, which stems from difficult foveal words and aims to lengthened current fixation for further processing.
An important issue here is the difference in their predictions to certain phenomena, one of which would be semantic preview benefit. In E-Z Reader model, L1 only accounts for a low-level process of the word. In order to access semantic
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
representation from word n+1, L1n+1 would have been done for L2n+1 to reach a certain level of completion. Under this circumstance, M1n+1 would have been canceled and a saccade to skip word n+2 is being planned, resulting in skipping of word n+1 without any detectable benefit on that word. Aside from mislocation of saccade, the only occasion for such preview benefit to appear is that by the time L1n+1 is done, saccade planning has reached M2n+1 (see Figure 2). The time for L2n+1 processing before the eyes saccade to word n+1 is then shorter than M2n+1, which is regarded as “formidable” by (Hohenstein & Kliegl, 2014). SWIFT model, on the other hand, does not exclude the possibility that a word whose semantic information has been activated becomes a potential target of saccade, since the selection is probability-based and is not only a function of activation level but also of other physical and lexical properties.
Figure 2. Semantic preprocessing predicted by E-Z Reader
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Another reason for the hot debate for semantic preview benefit is its elusiveness in English. As stated above, a certain type of preview benefit can be a writing-system-dependent phenomenon, such as morphological preview benefit for Hebrew. Semantic preview benefit, however, has been disputed in English, the most-studied writing system. The following section will review the studies that provide evidence, null or positive, for such effect in English and in other languages as well.