<?xml version="1.0" standalone="yes"?>
<Paper uid="P80-1043">
  <Title>Real Reading Behavior</Title>
  <Section position="3" start_page="160" end_page="160" type="metho">
    <SectionTitle>
((IWORD :HAS IDETERHINER-TAIL)
(DETERMINER-TAIL :HAS IWORO-EXPECTATION)
(WORD-EXPECTATION :IS IIWORD)
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="4" start_page="160" end_page="160" type="metho">
    <SectionTitle>
(WORD-EXPECTATION :IS (&lt;NEXTTOK&gt; WORD)))
</SectionTitle>
    <Paragraph position="0"> The number prefixes, as in &amp;quot;1WORD&amp;quot;, are tokens local to the production that just serve to indicate different knowledge base tokens are sought not what their knowledge base tokens should be. This production says that if a word has a determiner tail expecting some word and that word has been observed to be an adjective, then bring the confidence at least to 0.0 that the word-expectation is the adjective, and have confidence that the word-expectation is the word following the adjective.</Paragraph>
    <Paragraph position="1"> The &lt;SPEW&gt; action of this production makes use of a weighting scheme which serves to alter the control of processing. In this framework any knowledge base element can serve as both a bit of knowledge (a link) and as a control value. The .1 number causes the confidence in the source of the spew to be multiplied by -1 before it is added to the target, (WORD-EXPECTATION :IS 1WORD). If this were the only production requesting this switch of confidence, the effect would be the effective deletion of this bit of knowledge from the knowledge base. If other productions were also switching this confidence, the system would wind up being confident that this word-expectation association is indeed not the case (explicitly false).</Paragraph>
    <Paragraph position="2"> Processes in Sequence The primary interest in formulating a model is in having as much 'processing' or decision-making as possible in a single recognition-act cycle. The general idea is that an average gaze duration of 250 milliseconds on a word represents few such cycles. The ability of the model to predict gaze duration, then, depends upon the sequential constraints holding among the collection of productions brought to the interpretation process. The 'determiner tail' productions illustrated above represent a processing sequence in most contexts; the second cannot fire until the first has deposited its contribution in the knowledge base.</Paragraph>
    <Paragraph position="3"> This is not a necessary feature of these two productions, since other productions can collaborate to cause the simultaneous matching of the two productions illustrated (we assume these are easy to imagine). However, one may note that since the 'determiner tail' productions are distributed over several word gazes, they at most contribute one processing cycle to the gaze on any word (besides the determiner). Thus, sequencing over words may not be expensive. Let us consider where it is computationally expensive.</Paragraph>
    <Paragraph position="4"> In contrast to rvghtward looking activities, the presence of strong sequencing constraints among productions is potentially costly in leftward looking activities. To illustrate how such costs might be reduced, consider a production with a fairly low threshold which assigns a need to find an agent for an action-process verb, and another production which says that if one has an animate noun preceding an action-process verb and that animate noun is the only possible candidate, then that animate noun is the agent.</Paragraph>
    <Paragraph position="5"> These two productions are likely to fire simultaneously if the latter one fires at all. They both create a need to find an agent and satisfy that need at once. They do not set word * expectations simply because the look-back at previous text tries to be efficient with regard to sequencing constraints. Had the need not been immediately fulfilled, it would serve as a promotion of other productions which might find other ways of fulfilling it, or of reinterpreting the use of the action-process verb (even questioning the ISA inference). It should be noted that the natural device for keeping these further productions in sequence from firing is having them make the absence test, as in</Paragraph>
  </Section>
  <Section position="5" start_page="160" end_page="161" type="metho">
    <SectionTitle>
((!WORD :IS IACTION-PROCESS-VERB)
(WORD :HAS \]AGENT)
(&lt;ABSENT&gt; (AGENT :IS \]ANYTHING))
</SectionTitle>
    <Paragraph position="0"> --&gt; ...suggest this might be an imperative, passive, el\] ipse, etc.) The interpretation of the production is that &amp;quot;if you know with confidence that you have an action-process-verb and it needs an agent, but you don't know what that agent is, then suggest various reasons why you might not know with appropriately low confidence in them.&amp;quot;  Coordination of Mind and Eye The basic method of coordinating eye and mind in the present model is to make getting the next word contingent upon having completed the processing on the present one. In a production system architecture, this simply means that the match fails to turn up any productions whose conditions match to the knowledge base. Since elements in the knowledge base specify the need-to-know as wel: as what is known, the use of absence tests in the conditions of productions can 'shut off' further processing when it is deemed to be completed, or simply deemed to be unnecessary. It is by this device that the system demonstrates more processing on important information, 'shutting off' extended processing on that which is deemed, for any number of reasons, as less important.</Paragraph>
    <Paragraph position="1"> The model must, in addition to various ideas about coordination, be also capable of representing various ideas about dis-coordination. One potential instance of this in the present data is that while virtually every word is fixated upon at least once (recall that several fixations can count toward a single gaze), there are some words, AND, OR, BUT, A, THE, TO, and OF, with some likelihood of not being gazed upon at all (this accounts in some part for the fairly low average gaze duration on these words). This can be considered a dis-coordination of sorts, since to be this selective the reader must have some reasonable strong hypotheses about the words in question (the knowledge sources for these hypOtheses are potentially quite numerous, including the possibility of knowledge from peripheral vision). A production to implement this dis-coordination in the present  This production detects the presence of one of the above function words, and immediately shifts the present goal of interpreting a word (if it happens to be that) to gazing upon the word following the function word. It is important to recognize that the eye need not be on the function word for the system to know with reasonable confidence that the next word is a function word. The indexing scheme permits the system to form hypotheses strong enough to create effective reality (e.g., peripheral information and expectations can add up to the conclusion that the word is a function word). A second important property is that the system does not get confused with such skips, or in the usual case with such brief stays on these words. The reason again is because each word becomes a sort of local demon inheriting demon-like properties from general production, and by interaction with other knowledge base elements through the system of productions.</Paragraph>
    <Paragraph position="2"> Summary This report has provided a brief description on work in progress to capture our observations of reading eye-movements in computational models of the reading process. We have illustrated some of the main properties of reading eye-movements and some of the main issues to arise. We have also illustrated within an implemented system how these issues might be addressed and explored in order to gain insight into more precise queries about real reading behavior.</Paragraph>
    <Paragraph position="3"> Appendix An example text: Flywheels are one of the oldest mechanical devices known to man. Every internal-combustion engine contains a small flywheel that converts the jerky motion of the piston into the smooth flow of energy that powers the drive shaft. The greater the mass of a flywheel and the faster it spins, the more energy can be stored in it. But its maximum spinning speed is limited by the strength of the material it is made from. If it spins too fast for its mass, any flywheel will fly apart. One type of flywheel consists of round sandwiches of fiberglas and rubber providing the maximum possible storage of energy when the wheel is confined in a small space as in an automobile. Another type, the &amp;quot;superflywheel&amp;quot;, consists of a series of rimless spokes. This flywheel stores the maximum energy when space is unlimited.</Paragraph>
  </Section>
class="xml-element"></Paper>