The earliest psychology studies of reading can be traced back to the late 1800s. In 1879, Javal (as cited in [#!read-ed:singer2!#]) was the first researcher to publish data which revealed that reader's eyes do not move linearly and smoothly across a printed page; rather, the eyes jump in movements he dubbed saccades. This work led to a better understanding of when processing of text was taking place (during eye fixations) as well as proposing numerous questions as to how the processing could be improved. Following Javal's work, Cattell (as cited in [#!read-ed:singer2!#]) performed research which showed that proficient adult readers could perceive entire words at a rate faster than perception of its letters. This was an example of the automaticity which advanced readers achieve through practice. Other research of similar scope continued for the next several decades.
The first attempt to describe a more complete view of reading which went beyond the perceptual level did not occur until the 1950s. It was then that Holmes (read-ed:holmes1) proposed his substrata factor theory. This seminal work stated that reading speed and power was a result of a complex interaction between a large range of skills and cognitive processes. No single cognitive skill existed which could be tagged as ``the reading skill.'' While this conclusion seems obvious with the benefit of hindsight, it was a landmark result. It is also directly reflected in my own work--rather than developing the ``reading'' task in my theory, I have proposed a set of tasks which interact to produce the external behavior observers call ``reading.'' Holmes' work helped to lead to the methodology of functional theories which my work develops as its end result.
Another result, in a historical context, of the substrata theory was that researchers could now propose theories which explained pieces of the global puzzle. If it is accepted that reading contains numerous interacting tasks, it should be possible to develop theories of these individually. This will be the case unless the tasks are completely interdependent; however, if this was the case, there would be no reason to view the skill as being made up of the smaller tasks--part of the benefit of this is the separability.
One aspect which was studied was story grammar theory (e.g., [#!story-gram:mandler1!#,#!acm:sgram1!#]) which explains reading via a grammar representation--stories are coherent text passages with characters, a setting, and a series of actions (the plot description). All aspects of a story are not represented in a story grammar, a fact which they have been critiqued for. However, viewed from the substrata perspective this would be acceptable, if the researchers went on to examine the interactions which this set of theories has with the global reading activity. This line of research directly surfaces in my supertask theory in the role of the story structure comprehension supertask. However, I have been guided more by the overall framework in which my work is placed; as a result, my story structure comprehension supertask does not attempt to be all-powerful. It relies on the other supertasks to enable it to gather information relevant to the story structure. Two examples: the plot description is best derived by appealing to the scenario comprehension supertask, thereby lessening the role which story structure needs to play; characters can be identified by requesting information from the scenario comprehension supertask concerning the agents which it has discovered as acting in the scenario, as these will often be the characters of the story. As story grammar research and other psychological research into the nature of text structure and its influence on reading ability (e.g., [#!structure:meyer-1985!#]) have shown, human readers are able to make use of this structure. A theory of reading should take this into account as mine has been able to do. Chapter 5 will also describe a functional argument for the inclusion of this aspect of reading; this prior research also suggests the need for it.
Other theories of reading focused less on the structure of a piece of text and more on understanding the actions being performed by the characters. For example, many plan recognition theories were developed. Levy (read:levy), for example, believed that this identification of characters goals, beliefs, and plans is the crucial aspect of reading which needs to be studied. There were other attempts to use the idea of plan recognition as the basis for a reading comprehension system as well. The idea of a text communicating the intentions and plans of the characters involved can be traced back to the philosophical work of Grice (read:grice). His thesis was that to understand language (or text) completely, one must not only understand what the words mean but what the beliefs and intentions of the speaker (or characters) are. Prior to the work which attempted to make tenable systems based on this, there was a great deal of research dealing with planning in psychology and AI. By making use of this, some models attempted to directly understand text by identifying the plans of the characters. Plan recognition is a powerful technique, particularly when the reader is able to understand the interactions of characters and their plans, rather than understanding the plan of a single character. Still, it is too limiting to be a complete model of reading; a system concentrating on just plan recognition will fail to be a complete reading system.
This last set of research can be seen in my work under the scenario comprehension supertask. Each theory just described viewed the ``core'' of reading to be the comprehension of the events being described; these events could best be comprehended by understanding the goals and plans of the characters. As with the story structure comprehender, however, it is possible to place too much emphasis on this aspect of reading. Texts are not simply descriptions of events--they possess structure governed by a set of rules which can often be known a priori. Concentrating solely on the scenario will result in a model which is not taking advantage of all the information provided, thereby leading to too much effort being required.
Previous research often concentrated too heavily on one particular facet of reading, assuming that it would be sufficient to produce a theory. My framework approach leverages off these attempts as it integrates a wide variety of prior research, a task which often required me to extend or adapt the previous theories for my own research. The ``big'' picture was captured by the reading educators and their research; the specific pieces were often indicated by the psychological research being overviewed here and the artificial intelligence research seen in the next section. Without the interdisciplinary nature of my own work, these realizations would not have been made.
With that in mind, there are two more theories which I wish to present as they both attempt to capture this big picture. The first is the result of extensive research by Kintsch and van Dijk (read:kintsch). Their model proposes a unified process which is responsible for all of the characteristics of reading described in the above attempts. Text is transformed into propositions (units of information) which are combined by the reader into higher level understanding structures. The model has been extremely successful in modeling psychological data regarding the reading process, particularly those research areas dealing with recall of read material. Additionally, the model has been highly productive and has been updated and expanded by a number of researchers, including the original creators ([#!read:vandijk!#]). This construction-integration theory of reading is similar in scope to the ISAAC theory but differs greatly in its execution. First, the ISAAC theory is more modular with the cognitive ``load'' of reading distributed over more independent processes (the supertasks) than construction-integration uses. The previous history of reading indicates some good reasons for division of labor; I also have functional considerations which led me to that conclusion--to produce the three representations which were needed and to support the creative understanding process, I needed to have the six supertasks I developed. Second, in a related fashion, the Kintsch and van Dijk theory does not recognize the benefits of appealing to the story structure which texts possess. Third, there is less control at a meta-level of reading in their work, unlike my research which allows both reasoner modification of their reading dictated by their performance and their cognitive load, as well as reflection on the reading experience as an important aspect of the overall reading event (as captured by one of the three output representations, see Chapter 2). Finally, little is said of how the understanding of novel concepts would occur within this model of reading. Each of these points represents a difference in our philosophies of what is important for understanding the reading process and until further work can be done, especially in direct comparison of the two approaches, the relative strengths of each must remain somewhat unclear.
A similar result arises after viewing the work of Carpenter and Just (read:just). They chose to concentrate on the working memory constraints of the reader and how these influence reading and recall issues. They, too, have had a great deal of success modeling human subjects' performance within their READER framework. The primary focus of their work has always been on memory's influence on the reading process; in that regard, their work has had tremendous success and impact on future work. However, the use of memory in the Carpenter and Just theory is resource-bounded; in my model, memory only exists as long-term memory. This was sufficient for my work; in fact, my handling of long-term memory is more theoretically complete than the provisions made for it in the READER model. This is a result of different area of focus--they are interested in the low-level processes which are most closely tied to working memory (such as eye movement, control of information entering the reasoning system, and so forth) while I am interested in exploring higher-level processing, particularly what is involved in the understanding of novel concepts. Still, the READER model is important as it is a large-scale model which has had tremendous influence on later research, including being an inspiration to my own work.
In conclusion, many of the theories presented in this section concentrated on one aspect of the complete reading process, although the researchers acknowledged that a greater scope of reading behavior existed. Unfortunately, the theories which resulted often tried to accomplish more than they needed to if they had remained in the smaller focus area. The ISAAC theory attempts to keep each supertask as focused as possible, relying on the other supertasks to support each in the aspects of reading it may be weak in. This allows me to cover more of the area I concentrate on--creative understanding and how it affects reading. On the other hand, there have been psychological models which attempted to capture the full range of reading behavior. The output of these models is generally very good, as is the output of the ISAAC model. Each theory may, then, prove to be a sufficient one for describing the type of reading being focused on, as each theory does have a different view of the core element of reading. The ISAAC work uses a different set of techniques to accomplish what it does, a set which I believe to be sufficient for capturing a large amount of the creative reading experience, which neither of these prior models attempted. Only more research comparing the ``complete'' theories will be able to demonstrate the full range of differences between the various theories and make conclusions about them.