<?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="3" start_page="0" end_page="417" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Over the last two decades, the psycholinguistic literature has provided a wealth of experimental evidence for syntactic priming, i.e., the tendency to repeat syntactic structures (e.g., Bock, 1986).</Paragraph>
    <Paragraph position="1"> Most work on syntactic priming has been concerned with sentence production; however, recent studies also demonstrate a preference for structural repetition in human parsing. This includes the so-called parallelism effect demonstrated by Frazier et al. (2000): speakers processes coordinated structures more quickly when the second conjunct repeats the syntactic structure of the rst conjunct.</Paragraph>
    <Paragraph position="2"> Two alternative accounts of the parallelism effect have been proposed. Dubey et al. (2005) argue that the effect is simply an instance of a pervasive syntactic priming mechanism in human parsing. They provide evidence from a series of corpus studies which show that parallelism is not limited to co-ordination, but occurs in a wide range of syntactic structures, both within and between sentences, as predicted if a general priming mechanism is assumed. (They also show this effect is stronger in coordinate structures, which could explain Frazier et al.'s (2000) results.) Frazier and Clifton (2001) propose an alternative account of the parallelism effect in terms of a copying mechanism. Unlike priming, this mechanism is highly specialized and only applies to co-ordinate structures: if the second conjunct is encountered, then instead of building new structure, the language processor simply copies the structure of the rst conjunct; this explains why a speed-up is observed if the two conjuncts are parallel. If the copying account is correct, then we would expect parallelism effects to be restricted to coordinate structures and not to apply in other contexts. This paper presents a parsing model which implements both the priming mechanism and the copying mechanism, making it possible to compare their predictions on human reading time data.</Paragraph>
    <Paragraph position="3"> Our model also simulates other important aspects of human parsing: (i) it is broad-coverage, i.e., it yields accurate parses for unrestricted input, and (ii) it processes sentences incrementally, i.e., on a word-by-word basis. This general modeling framework builds on probabilistic accounts of human parsing as proposed by Jurafsky (1996) and Crocker and Brants (2000).</Paragraph>
    <Paragraph position="4"> A priming-based parser is also interesting from an engineering point of view. To avoid sparse data problems, probabilistic parsing models make strong independence assumptions; in particular, they generally assume that sentences are independent of each other, in spite of corpus evidence for structural repetition between sentences. We therefore expect a parsing model that includes structural repetition to provide a better t with real corpus data, resulting in better parsing performance.</Paragraph>
    <Paragraph position="5"> A simple and principled approach to handling structure re-use would be to use adaptation probabilities for probabilistic grammar rules (Church, 2000), analogous to cache probabilities used in caching language models (Kuhn and de Mori, 1990). This is the approach we will pursue in this paper.</Paragraph>
    <Paragraph position="6"> Dubey et al. (2005) present a corpus study that demonstrates the existence of parallelism in corpus data. This is an important precondition for understanding the parallelism effect; however, they  do not develop a parsing model that accounts for the effect, which means they are unable to evaluate their claims against experimental data. The present paper overcomes this limitation. In Section 2, we present a formalization of the priming and copying models of parallelism and integrate them into an incremental probabilistic parser. In Section 3, we evaluate this parser against reading time data taken from Frazier et al.'s (2000) parallelism experiments. In Section 4, we test the engineering aspects of our model by demonstrating that a small increase in parsing accuracy can be obtained with a parallelism-based model. Section 5 provides an analysis of the performance of our model, focusing on the role of the distance between prime and target.</Paragraph>
  </Section>
class="xml-element"></Paper>