<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0313">
  <Title>Modeling sentence processing in ACT-R</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 A brief introduction to the cognitive
architecture ACT-R
</SectionTitle>
    <Paragraph position="0"> ACT-R is a theory of the human cognitive architecture. It allows the development of computational models that can closely simulate experimental methodologies such as eye-tracking and self-paced reading, and has been used to model a wide array of behavioral data from learning and memory, problem solving and decision making, language and communication, perception and attention, cognitive development, and individual differences (Anderson et al., 2002).</Paragraph>
    <Paragraph position="1"> The ACT-R architecture is attractive as a modeling tool for three reasons. First, it is based on a wide array of empirical results in various domains of cognitive psychology. Second, it is flexible enough to permit the modeler to add their own assumptions and theories about the specific task to be modeled.</Paragraph>
    <Paragraph position="2"> Finally, ACT-R models yield dependent measures such as reading time in much the same way as humans performing the experiment; e.g., the system can easily be programmed to simulate key presses after it processes material presented on the screen.</Paragraph>
    <Paragraph position="3"> As shown in Figure 1, the architecture consists of several MODULES such as Declarative, Visual, and Manual. Each module is associated with a BUFFER which temporarily stores information for a given action. For example, the visual buffer is used to store an item &amp;quot;seen&amp;quot; by the system in the environment before it is used in the service of some task.</Paragraph>
    <Paragraph position="4"> The module that is especially important for the present paper is the Declarative (henceforth, DM).</Paragraph>
    <Paragraph position="5"> DM represents permanent memory: every fact that is assumed to be known is encoded as a CHUNK in declarative memory. A chunk is an attribute-value list structure with a special attribute, ISA, which defines its type. The attributes are also referred to as slots. The value of a chunk's slot is also (by definition) a chunk, unless it is double-quoted or is the  system. &amp;quot;Environment&amp;quot; is the outside world that ACT-R is programmed to interact with. The arrows show the possible flows of information. Productions and the central box with the boxes labeled &amp;quot;Matching&amp;quot;, &amp;quot;Selection&amp;quot;, and &amp;quot;Execution&amp;quot; are intended to represent a set of central executive mechanisms and processes.</Paragraph>
    <Paragraph position="6"> lisp primitive &amp;quot;nil&amp;quot;.</Paragraph>
    <Paragraph position="7"> Each DM chunk has an activation that determines its speed of retrieval, and the probability that it will be retrieved; the initial activation for a given chunk can be set manually.</Paragraph>
    <Paragraph position="8"> There is a GOAL BUFFER that holds a current goal under consideration (there can be only one goal at one time); this goal is a chunk with a given type and possibly instantiated slots.</Paragraph>
    <Paragraph position="9"> The control structure for modeling a sequence of events is a set of PRODUCTIONS; a production is simply an if-then statement of the following general form: for a given state of one or more buffers and/or DM, execute some actions. Examples of executing actions are retrieving something from DM; changing a value in one of the goal's slots; repositioning the hand over a keyboard; a visual shift of attention; changing the goal to a new one, etc. If the goal is changed, then this new goal now occupies the goal buffer.</Paragraph>
    <Paragraph position="10"> Building an ACT-R model is essentially a definition of possible sequences of actions for a given state of affairs. Events like retrievals from DM are triggered by looking at the contents of one or more buffers. For example, the ACT-R system &amp;quot;sees&amp;quot; an item/object on the screen and then encodes it as a visual chunk. This chunk can then be harvested from the visual buffer; it includes (as slot-value specifications) information about the content of the item seen, its x-y coordinates, etc. One can define an action based on this information, such as retrieving a chunk from DM.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="300" type="metho">
    <SectionTitle>
3 Modeling sentence parsing in ACT-R
</SectionTitle>
    <Paragraph position="0"> Previous research suggests that humans employ some variant of left-corner parsing (see, e.g., (Resnik, 1992)), which in essence involves a bottom-up and a top-down (predictive) step. We adopt this parsing strategy in the simulations. In order to model the prediction of syntactic structure based on incrementally appearing input, we assume that sentence structure templates are available in declarative memory as underspecified chunks.</Paragraph>
    <Paragraph position="1"> These chunks are retrieved every time a new word is integrated into the structure, as are prior arguments necessary for semantic integration.</Paragraph>
    <Paragraph position="2"> We illustrate the parsing process with a simple example (Figure 2). Suppose that the sentence to be parsed is The girl ran, and suppose that we are simulating self-paced reading (Just et al., 1982). When the word the is seen, a bottom-up and top-down structure building step results in a sentence with an intransitive verb being predicted. This structure becomes the current goal. Then the word girl is seen and processed, i.e., its lexical entry is retrieved from declarative memory. The noun slot in the goal is then instantiated with that lexical entry. In the next step, if the word ran is seen the relevant lexical item for the verb is retrieved and instantiated with the verb slot of the goal; here, the verb's argument is also retrieved and integrated with the subcategorization frame of the verb. If, instead of ran the word that appears, a new goal is created, with any previously instantiated slots of the preceding goal being passed on to the new goal, and parsing proceeds from there.</Paragraph>
    <Paragraph position="3"> Each retrieval of a goal from memory results in a surge in its activation, so that repeated retrievals result in increased activation; and the higher the activation of an item the faster it is processed. At the same time, activation decays according to the power law of forgetting (Anderson et al., 2002). In the same way that the goals undergo decay and reactivation, so do the previously seen words. This means that the speed of retrieval of a previously seen argument at a verb will be determined by the activation level of that argument. Thus, the activation of both the goals (predicted structures) and the arguments affect processing.</Paragraph>
    <Paragraph position="4"> In our simulations, for simplicity we code in the exact steps that ACT-R takes for particular sentences. Although it is feasible to build a very gen- null eral parser in pure ACT-R, before doing this we wanted to first establish whether ACT-R's reactivation mechanisms can account for a reasonable array of facts from the sentence processing literature. In (Lewis and Vasishth, An activation-based model of sentence processing as skilled memory retrieval, (tentative title; in preparation)) we provide a detailed description of a model employing mechanisms similar to those described here, but one that behaves more like a standard parser.</Paragraph>
    <Section position="1" start_page="0" end_page="300" type="sub_section">
      <SectionTitle>
3.1 English subject versus object relative
clauses
</SectionTitle>
      <Paragraph position="0"> It is well known (Homes and O'Regan, 1981) that English subject relatives are easier to process that object relatives (1). In the parsing model outlined above, we can model this result without changing  any ACT-R parameters at all (i.e., we use the default settings for the parameters).</Paragraph>
      <Paragraph position="1"> (1) a. The reporter who sent the photographer  to the editor hoped for a good story.</Paragraph>
      <Paragraph position="2"> b. The reporter who the photographer sent to the editor hoped for a good story.</Paragraph>
      <Paragraph position="3"> The explanation comes from the decay of the arguments of the verb sent: in object relatives the argument reporter decays much more than in the subject relative by the time it is integrated with the verb's subcategorization frame (Figure 3). This is because more time elapses between the argument being first seen and its retrieval at the verb.1 1A reviewer points out that several head-final languages such as German and Dutch also have a subject relative preference and in these languages the activation level cannot be the explanation. We do not claim that decay is the only constraint operating in parsing; frequency effects (greater preference for  Retrieval of reporter at sent is harder in the object relative because of increased argument decay.</Paragraph>
    </Section>
    <Section position="2" start_page="300" end_page="300" type="sub_section">
      <SectionTitle>
3.2 The SC/RC asymmetry in English
</SectionTitle>
      <Paragraph position="0"> It is also well-known (Gibson, 2000) that a sentential complement (SC) followed by a relative clause  (RC) is easier to process than an RC followed by an SC: (2) a. The fact that the employee who the manager hired stole office supplies worried the executive.</Paragraph>
      <Paragraph position="1"> b. #The executive who the fact that the  employee stole office supplies worried hired the manager.</Paragraph>
      <Paragraph position="2"> As in the previous discussion about relative clauses, in the harder case the decay of the argument executive at the verb worried is greater compared to the decay of the argument employee at hired in the easier-to-process sentence. In addition, the total reading time for the harder sentence is about 120 msec longer.2</Paragraph>
    </Section>
    <Section position="3" start_page="300" end_page="300" type="sub_section">
      <SectionTitle>
3.3 Hindi center embeddings
</SectionTitle>
      <Paragraph position="0"> Previous work (Hakes, 1972), (Konieczny, 2000) has shown that if argument-verb distance is increased, processing is easier at the verb. (Vasishth, more frequently occurring subject relatives) etc. could certainly dominate where the amount of decay is constant in subject and object relatives. It is an open empirical question whether frequency alone can account for the subject/object asymmetry in English, but given that we have independent empirical justification for decay (see Section 3.5), the above is a plausible explanation.</Paragraph>
      <Paragraph position="1"> 2As a reviewer points out, &amp;quot;the account in terms of activation decay suggests that the SC/RC asymmetry can be annihilated or even reversed by inserting longer or shorter NPs between the critical verbs (worried, hired) and their arguments (executive, employee). This seems unrealistic.&amp;quot; This is surely an empirical question that needs to be verified experimentally; we intend to pursue this very interesting issue in future work.</Paragraph>
      <Paragraph position="2">  2003) presented similar results in Hindi. The Hindi experiment manipulated distance by comparing the baseline condition (3a) with the case where an adverb intervened (3b), a verb-modifying PP intervened (3c), and relative clause intervened that modified the preceding NP (3d).</Paragraph>
      <Paragraph position="3">  'Sita told Hari to tell Ravi to buy the book that was lying on a/the table.' In all the &amp;quot;insertion&amp;quot; cases a statistically significant speedup was observed at the verb, compared to the baseline condition.</Paragraph>
      <Paragraph position="4"> This experiment's results were replicated in the ACT-R system; the replication is based on the assumption that the goal (predicted syntactic structure) is reactivated each time it (i.e., the entire predicted structure) is modified. The intervening items result in an extra retrieval compared to the baseline, resulting in faster processing at the verb. In this model, one parameter was changed: the rate of decay of items. We justify this change in the next sub-section.</Paragraph>
      <Paragraph position="5"> The modeling results are shown in Figure 5.</Paragraph>
    </Section>
    <Section position="4" start_page="300" end_page="300" type="sub_section">
      <SectionTitle>
3.4 Individual variation in Hindi center
</SectionTitle>
      <Paragraph position="0"> embedding data In the Hindi experiment, there was a further variation in the data when individual subjects' data were considered: only about 48% of subjects showed a speedup at the verb. About 21% showed a slowdown and there was only a few milliseconds difference (essentially no difference) in the reading times for about 31% of the subjects. The observed variation was a systematic trend in the sense that the 47% of the subjects who showed a speedup or slowdown in adverb-insertion case also showed the same trend in the PP- and RC-inserted cases - the probability of this happening is considerably below chance level.</Paragraph>
      <Paragraph position="1"> The rate of decay defined in ACT-R's rehearsal equation can systematically explain this variation.</Paragraph>
      <Paragraph position="2"> Consider the situation where a chunk a0 with an initial activation of a1a3a2 is retrieved. The activation is  recalculated each time a retrieval occurs, according to the following equation.</Paragraph>
      <Paragraph position="4"> Here, a23 is the number of times the chunk a0 was successfully retrieved,  a10 is the time elapsed since the a24 -th retrieval, and a25 is a decay rate that defaults to a26a28a27a30a29 in ACT-R. This equation reflects the log odds that a chunk would reoccur as function of how it has appeared in the past (Anderson et al., 2002, 17). It turns out that the a25 parameter take us beyond boolean predictions: a25 a0 a27a31a26a33a32 results in a speedup; a25 a0 a27a30a29 results in a slowdown; and a25 a0 a27a4a32a35a34 results in no difference in RT at the verb; see Figures 6 to 8.3</Paragraph>
    </Section>
    <Section position="5" start_page="300" end_page="300" type="sub_section">
      <SectionTitle>
3.5 Comparison with other models
</SectionTitle>
      <Paragraph position="0"> The model presented here is very different in conception from existing models of sentence processing. For example, consider Early Immediate Consistuents (Hawkins, 1994) and Discourse Locality Theory (Gibson, 2000), two theories with significant empirical coverage. Both theories propose variants of what we will call the distance hypothesis: increasing the distance between arguments and a subsequently appearing verb (head) that selects for them results in increased processing difficulty at the verb. Distance here is quantified in terms of the number of words in a constituent (EIC) or the number of new discourse referents introduced between the arguments and head (DLT).</Paragraph>
      <Paragraph position="1"> The present model claims that distance effects are actually a result of argument decay. Preliminary evidence that it is really decay and not EICor DLT-defined distance comes from a recent self-paced listening experiment (Vasishth et al., 2004) in which two conditions were contrasted: arguments and verbs with (a) an adjunct intervening, (b) si- null In (5), the arguments kaagaz, 'paper', and lar. kaa, 'boy' are separated from the verb dekhaa, 'saw' by 3Of course, modeling individual variation in terms of differing rates of decay assumes that subjects exhibit varying degrees of decay rates. An experiment is currently in progress that attempts to correlate varying verbal sentence span with subject behavior in the insertion cases.</Paragraph>
      <Paragraph position="2"> an adjunct containing two4 discourse referents (5a); or by silence (5b). Subjects were allowed to interrupt the silence and continue listening to the rest of the sentence whenever they wanted to. Subjects interruped the silence (on an average) after about 1.4 seconds.</Paragraph>
      <Paragraph position="3"> Distance based theories predict that having an intervening adjunct that introduces discourse referents should result in greater processing difficulty at the verb dekhaa, 'saw', compared to when silence intervenes. If decay rather than distance is the critical factor here that affects processing, then there should be greater difficulty at the verb in the silence condition than when in the items intervene (see Section 3.3 for why intervening items may facilitate processing). The results support the activation account: introducing silence results in significantly longer reading times at the verb dekhaa than when intervening items occur.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>