<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1053">
  <Title>Integrating Syntactic Priming into an Incremental Probabilistic Parser, with an Application to Psycholinguistic Modeling</Title>
  <Section position="9" start_page="423" end_page="423" type="concl">
    <SectionTitle>
6 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> The main contribution of this paper has been to show that an incremental parser can simulate syntactic priming effects in human parsing by incorporating probability models that take account of previous rule use. Frazier et al. (2000) argued that the best account of their observed parallelism advantage was a model in which structure is copied from one coordinate sister to another. Here, we explored a probabilistic variant of the copy mechanism, along with two more general models based on within- and between-sentence priming. Although the copy mechanism provided the strongest parallelism effect in simulating the human reading time data, the effect was also successfully simulated by a general within-sentence priming model.</Paragraph>
    <Paragraph position="1"> On the basis of simplicity, we therefore argue that it is preferable to assume a simpler and more general mechanism, and that the copy mechanism is not needed. This conclusion is strengthened when we turn to consider the performance of the parser on the standard Penn Treebank test set: the Within model showed a small increase in F-score over the PCFG baseline, while the copy model showed no such advantage.5 All the models we proposed offer a broad-coverage account of human parsing, not just a limited model on a hand-selected set of examples, such as the models proposed by Jurafsky (1996) and Hale (2001) (but see Crocker and Brants 2000).</Paragraph>
    <Paragraph position="2"> A further contribution of the present paper has been to develop a methodology for analyzing the (re-)use of syntactic rules over time in a corpus. In particular, we have de ned an algorithm for randomizing the constituents of a treebank, yielding a baseline estimate of chance repetition.</Paragraph>
    <Paragraph position="3"> In the research reported in this paper, we have adopted a very simple model based on an unlexicalized PCFG. In the future, we intend to explore the consequences of introducing lexicalization into the parser. This is particularly interesting from the point of view of psycholinguistic modeling, because there are well known interactions between lexical repetition and syntactic priming, which require lexicalization for a proper treatment. Future work will also involve the use of smoothing to increase the bene t of priming for parsing accuracy. The investigations reported 5The broad-coverage parsing experiment speaks against a 'facilitation' hypothesis, i.e., that the copying and priming mechanisms work together. However, a full test of this (e.g., by combining the two models) is left to future research. in Section 5 provide a basis for estimating the smoothing parameters.</Paragraph>
  </Section>
class="xml-element"></Paper>