<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-3007">
  <Title>The Acquisition of Stress: A Data-Oriented Approach</Title>
  <Section position="6" start_page="447" end_page="448" type="concl">
    <SectionTitle>
5. Conclusion
</SectionTitle>
    <Paragraph position="0"> Dresher and Kaye (1990:146) argue that &amp;quot;A rich and highly structured theory of UG \[Universal Grammar\] is otiose if the same results can be achieved by simpler means.&amp;quot; What might these alternatives be? A possible alternative is a data-oriented one, which can be described as follows: it appears that stress patterns are sensitive to sequences of syllables and syllable weights. We could simply map strings of weighted syllables (weight strings) into sequences of stresses (stress strings). In this way a record would be kept of the stress strings associated with each weight string. This alternative was suggested by Church (1992) 13 and by Dresher and Kaye (Dresher and Kaye 1990; Dresher 1992). Dresher (1992) concludes the discussion of this alternative by stating that it is empirically inadequate: &amp;quot;It would be unable to project its grammar to assign stress to weight strings not yet encountered&amp;quot; (Dresher 1992: 301).</Paragraph>
    <Paragraph position="1"> In this paper we investigated a learning device that incorporates the very simple data-driven alternative described above. The memory of an Instance-Based Learning (IBL) system is a kind of table in which representations of words are associated with 13 Church (1992), in a reaction to Dresher (1992) proposes lookup in a table of syllable weight strings (associated with their stress string) as an alternative approach. However, he glosses over the problem of ambiguity and noise, and of how to arrive at a syllable weight representation on the basis of the spelling of words.</Paragraph>
    <Paragraph position="2">  Walter Daelemans, Gert Durieux, and Steven Gillis The Acquisition of Stress stress assignments. However, by using simple similarity-based reasoning, the algorithm can generalize beyond the data on which it was trained.</Paragraph>
    <Paragraph position="3"> We showed that IBL was able to acquire a considerable portion of the regularities governing the stress system of Dutch. This finding is in agreement with a similar enterprise undertaken by Gupta and Touretzky (1994), who used a simple perceptron as their data-driven approach, and shows that the tree-building operations proposed in learning theories for metrical phonology are not necessary for learning stress assignment. null We also investigated two other aspects of the learning process. First of all, in the learning experiments described in the literature thus far, only stereotyped representations of the stress patterns of languages have been used as learning material. In this study we used a lexicon of 4868 attested monomorphemes. This lexicon showed all the general characteristics of the intricate Dutch stress system, but it also contained a fair amount of noise: exceptional words from a metrical point of view as well as plain irregular cases. It was shown that IBL discovered the regularities despite the noise.</Paragraph>
    <Paragraph position="4"> Secondly, we investigated the effect of using different representations for the training material of the learner on the learning results. The input encodings reflected the amount of 'theoretical bias' or a priori knowledge that a learner could be provided with. More specifically, a weight-string encoding is considered to be necessary and sufficient in the literature for learning a quantity-sensitive language such as Dutch. We contrasted such an encoding of our learning material with an encoding that consisted of rhyme projections, and with a plain phonemic representation (that included syllable boundaries). It turned out that the phonemic representations yield significantly better results than the encoding in terms of syllable weights. This implies that a data-driven approach to the task of acquiring main stress assignment is feasible even without the a priori knowledge incorporated in weight-strings.</Paragraph>
    <Paragraph position="5"> Taken together, our results suggest that the representations and operations specified by metrical theory may be neither necessary nor sufficient for learning stress assignment. More specifically, information about segmental content may warrant more attention in metrical phonology. More generally, the results weaken Dresher and Kaye's (1990) argument for the necessity of a principles and parameters approach to the acquisition of stress.</Paragraph>
  </Section>
class="xml-element"></Paper>