<?xml version="1.0" standalone="yes"?>
<Paper uid="P96-1040">
  <Title>The Rhythm of Lexical Stress in Prose</Title>
  <Section position="2" start_page="0" end_page="302" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Rhythm inheres in creative output, asserting itself as the meter in music, the iambs and trochees of poetry, and the uniformity in distances between objects in art and architecture. More subtly there is widely believed to be rhythm in English prose, reflecting the arrangement of words, whether deliberate or subconscious, to enhance the perceived acoustic signal or reduce the burden of remembrance for the reader or author.</Paragraph>
    <Paragraph position="1"> In this paper we describe an information-theoretic model based on lexical stress that substantiates this common perception and relates stress regularity in written speech (which we shall equate with the intuitive notion of &amp;quot;rhythm&amp;quot;) to the probability of the text itself. By computing the stress entropy rate for both a set of Wall Street Journal sentences and a version of the corpus with randomized intra-sentential word order, we also find that word order contributes significantly to rhythm, particularly within highly probable sentences. We regard this as a first step in quantifying the extent to which metrical properties influence syntactic choice in writing.</Paragraph>
    <Section position="1" start_page="0" end_page="302" type="sub_section">
      <SectionTitle>
1.1 Basics
</SectionTitle>
      <Paragraph position="0"> In speech production, syllables are emitted as pulses of sound synchronized with movements of the musculature in the rib cage. Degrees of stress arise from variations in the amount of energy expended by the speaker to contract these muscles, and from other factors such as intonation. Perceptually stress is more abstractly defined, and it is often associated with &amp;quot;peaks of prominence&amp;quot; in some representation of the acoustic input signal (Ochsner, 1989).</Paragraph>
      <Paragraph position="1"> Stress as a lexical property, the primary concern of this paper, is a function that maps a word to a sequence of discrete levels of physical stress, approximating the relative emphasis given each syllable when the word is pronounced. Phonologists distinguish between three levels of lexical stress in English: primary, secondary, and what we shall call weak for lack of a better substitute for unstressed. For the purposes of this paper we shall regard stresses as symbols fused serially in time by the writer or speaker, with words acting as building blocks of pre-defined stress sequences that may be arranged arbitrarily but never broken apart.</Paragraph>
      <Paragraph position="2"> The culminative property of stress states that every content word has exactly one primary-stressed syllable, and that whatever syllables remain are subordinate to it. Monosyllabic function words such as the and of usually receive weak stress, while content words get one strong stress and possibly many secondary and weak stresses.</Paragraph>
      <Paragraph position="3"> It has been widely observed that strong and weak tend to alternate at &amp;quot;rhythmically ideal disyllabic distances&amp;quot; (Kager, 1989a). &amp;quot;Ideal&amp;quot; here is a complex function involving production, perception, and many unknowns. Our concern is not to pinpoint this ideal, nor to answer precisely why it is sought by speakers and writers, but to gauge to what extent it is sought.</Paragraph>
      <Paragraph position="4"> We seek to investigate, for example, whether the avoidance of primary stress clash, the placement of two or more strongly stressed syllables in succession, influences syntactic choice. In the Wall Street Jour- null nal corpus we find such sentences as &amp;quot;The fol-lowing is-sues re-cent-ly were filed with the Se-curi-ties and Ex-change Com-mis-sion&amp;quot;. The phrase &amp;quot;recently were filed&amp;quot; can be syntactically permuted as &amp;quot;were filed recently&amp;quot;, but this clashes filed with the first syllable of recently. The chosen sentence avoids consecutive primary stresses. Kager postulates with a decidedly information theoretic undertone that the resulting binary alternation is &amp;quot;simply the maximal degree of rhythmic organization compatible with the requirement that adjacent stresses are to be avoided.&amp;quot; (Kager, 1989a) Certainly we are not proposing that a hard decision based only on metrical properties of the output is made to resolve syntactic choice ambiguity, in the case above or in general. Clearly semantic emphasis has its say in the decision. But it is our belief that rhythm makes a nontrivial contribution, and that the tools of statistics and information theory will help us to estimate it formally. Words are the building blocks. How much do their selection (diction) and their arrangement (syntax) act to enhance rhythm?</Paragraph>
    </Section>
    <Section position="2" start_page="302" end_page="302" type="sub_section">
      <SectionTitle>
1.2 Past models and quantifications
</SectionTitle>
      <Paragraph position="0"> Lexical stress is a well-studied subject at the intra-word level. Rules governing how to map a word's orthographic or phonetic transcription to a sequence of stress values have been searched for and studied from rules-based, statistical, and connectionist perspectives. null Word-external stress regularity has been denied this level of attention. Patterns in phrases and compound words have been studied by Halle (Halle and Vergnaud, 1987) and others, who observe and reformulate such phenomena as the emphasis of the penultimate constituent in a compound noun (National Center for Supercomputing Applications, for example.) Treatment of lexical stress across word boundaries is scarce in the literature, however. Though prose rhythm inquiry is more than a hundred years old (Ochsner, 1989), it has largely been dismissed by the linguistic community as irrelevant to formal models, as a mere curiosity for literary analysis. This is partly because formal methods of inquiry have failed to present a compelling case for the existence of regularity (Harding, 1976).</Paragraph>
      <Paragraph position="1"> Past attempts to quantify prose rhythm may be classified as perception-oriented or signal-oriented.</Paragraph>
      <Paragraph position="2"> In both cases the studies have typically focussed on regularities in the distance between peaks of prominence, or interstress intervals, either perceived by a human subject or measured in the signal. The former class of experiments relies on the subjective segmentation of utterances by a necessarily limited number of participants--subjects tapping out the rhythms they perceive in a waveform on a recording device, for example (Kager, 1989b). To say nothing of the psychoacoustic biases this methodology introduces, it relies on too little data for anything but a sterile set of means and variances.</Paragraph>
      <Paragraph position="3"> Signal analysis, too, has not yet been applied to very large speech corpora for the purpose of investigating prose rhythm, though the technology now exists to lend efficiency to such studies. The experiments have been of smaller scope and geared toward detecting isochrony, regularity in absolute time. Jassem et al.(Jassem, Hill, and Witten, 1984) use statistical techniques such as regression to analyze the duration of what they term rhythm units.</Paragraph>
      <Paragraph position="4"> Jassem postulates that speech is composed of extrasyllable narrow rhythm units with roughly fixed duration independent of the number of syllable constituents, surrounded by varia.ble-length anacruses. Abercrombie (Abercrombie, 1967) views speech as composed of metrical feet of variable length that begin with and are conceptually highlighted by a single stressed syllable.</Paragraph>
      <Paragraph position="5"> Many experiments lead to the common conclusion that English is stress-timed, that there is some regularity in the absolute duration between strong stress events. In contrast to postulated syllabletimed languages like French in which we find exactly the inverse effect, speakers of English tend to expand and to contract syllable streams so that the duration between bounding primary stresses matches the other intervals in the utterance. It is unpleasant for production and perception alike, however, when too many weak-stressed syllables are forced into such an interval, or when this amount of &amp;quot;padding&amp;quot; varies wildly from one interval to the next. Prose rhythm analysts so far have not considered the syllable stream independent from syllabic, phonemic, or interstress duration. In particular they haven't measured the regularity of the purely lexical stream.</Paragraph>
      <Paragraph position="6"> They have instead continually re-answered questions concerning isochrony.</Paragraph>
      <Paragraph position="7"> Given that speech can be divided into interstress units of roughly equal duration, we believe the more interesting question is whether a speaker or writer modifies his diction and syntax to fit a regular number of syllables into each unit. This question can only be answered by a lexical approach, an approach that pleasingly lends itself to efficient experimentation with very large amounts of data.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>