<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-0604">
  <Title>Probing the space of grammatical variation: induction of cross-lingual grammatical constraints from treebanks</Title>
  <Section position="8" start_page="26" end_page="27" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> Probabilistic language models, machine language learning algorithms and linguistic theorizing all appear to support a view of language processing as a process of dynamic, on-line resolution of conflicting grammatical constraints. We begin to gain considerable insights into the complex process of bootstrapping nature and behaviour of these constraints upon observing their actual distribution in perceptually salient contexts. In our view of things, this trend outlines a promising framework providing fresh support to usage-based models of language acquisition through mathematical and computational simulations. Moreover, it allows scholars to investigate patterns of cross-linguistic typological variation that crucially depend on the appropriate setting of model parameters. Finally, it promises to solve, on a principled basis, traditional performance-oriented cruces of grammar theorizing such as degrees of human acceptability of ill-formed grammatical constructions (Hayes 2000) and the inherently graded compositionality of linguistic constructions such as morpheme-based words and word-based phrases (Bybee 2002, Hay and Baayen 2005).</Paragraph>
    <Paragraph position="1"> We argue that the current availability of comparable, richly annotated corpora and of mathematical tools and models for corpus exploration make time ripe for probing the space of grammatical variation, both intra- and interlinguistically, on unprecedented levels of sophistication and granularity. All in all, we anticipate that such a convergence is likely to have a twofold impact: it is bound to shed light on the integration of performance and competence factors in language study; it will make mathematical models of language increasingly able to accommodate richer and richer language evidence, thus putting explanatory theoretical accounts to the test of a usage-based empirical verification.</Paragraph>
    <Paragraph position="2"> In the near future, we intend to pursue two parallel lines of development. First we would like to increase the context-sensitiveness of our processing task by integrating binary grammatical constraints into the broader context of multiply conflicting grammar relations. This way, we will be in a position to capture the constraint that a (transitive) verb has at most one subject and one object, thus avoiding multiple assignment of subject (object) relations in the same context. Suppose, for example, that both nouns in a noun-noun-verb triple are amenable to a subject interpretation, but that one of them is a more likely subject than the other. Then, it is reasonable to expect the model to process the less likely subject candidate as the object of the verb in the triple. Another promising line of development is based on the observation that the  order in which verb arguments appear in context is also lexically governed: in Italian, for example, report verbs show a strong tendency to select subjects post-verbally. Dell'Orletta et al. (2005) report a substantial improvement on the model performance on Italian SOI when lexical information is taken into account, as a lexicalized MaxEnt model appears to integrate general constructional and semantic biases with lexically-specific preferences. In a cross-lingual perspective, comparable evidence of lexical constraints on word order would allow us to discover language-wide invariants in the lexicon-grammar interplay.</Paragraph>
  </Section>
class="xml-element"></Paper>