<?xml version="1.0" standalone="yes"?>
<Paper uid="P86-1023">
  <Title>Morph~lo~leal Decomposition and 5tress Assignment for Speech Synthesis</Title>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
* Dictionary Lookup
* Letter to Sound
</SectionTitle>
    <Paragraph position="0"> Both appt~oaches have their advantages and disadvantages; dictionary lookup fails for unknown words (e.g., proper nouns) and letter to sound rules fail for irregular words, which are all too common in English. Most speech synthesizers adopt a hybrid strategy, using the dictionary when possible and turning to letter to sound rules for the rest. I discussed letter to sound rules at the last meeting of the ACL \[Church\]; this paper will report on some new dictionary lookup approaches, with an emphasis on morphology.</Paragraph>
    <Paragraph position="1"> Morphological decomposition is used to reduce the size of the dictionary and to increase coverage. Instead of storing all possible words, the system can store just a lexicon of morphemes and save a factor of 10 \[Jon Allen (personal communication)\] in storage. Now when the system is given a word and asked to determine is pronunciation, the system decomposes the word into known morphemes, looks up the pronunciation of each of the pieces and combines the results.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="156" type="metho">
    <SectionTitle>
2. MITalk Decomp
</SectionTitle>
    <Paragraph position="0"> The best known morphological decomposition system is the Decomp module in the MITalk sysnthesizer \[Allen et. al.\]. This system attempted to parse an input word such as formally into morphemes: form, -al and -ly. It was assumed that morphemes are concatenated together (like &amp;quot;beads on a string&amp;quot;) according to the finite state grammar shown below: The types of morphemes were:  a. Free (root): stay, squeeze, large b. Absolute (absl): the, than, but c. Left-Bound (lbrt): rePEL, conCEIVE d. Right-Bound (rbrt): CRIMINal, TOLERance e. Strong (root): women, rang  Costs were placed on the arcs to alleviate overgeneration. Note that the grammar produces quite a number of spurious analyses. For example, not only would formally be analyzed as form-al-ly but it would also be analyzed as form-ally and for-mal-ly. The cost mechanism blocks these spurious analyses by assigning compounding a higher cost than suffixation and therefore favoring the desired analysis. Although the cost mechanism handles a large number of cases, it would be better to aim toward a tighter grammar of morphology which did not overgenerate so badly.</Paragraph>
    <Section position="1" start_page="156" end_page="156" type="sub_section">
      <SectionTitle>
State Arc Cost
</SectionTitle>
      <Paragraph position="0"> word-final: cat infl word-final 64 cat derv right-sida-a 35 cat root left-side-a 101 cat lbrt middle 1091 cat absl word-initial 1221 right-side-a: cat derv right-side-a 35 cat infl word-final 35 cat rbrt left-side-a 66 cat root left-side-a 101 cat lbrt middle 1091 right-side-b: cat derv right-side-a 963 cat lbrt middle 2019 cat infl word-final 992 cat root left-side-a 1029 cat rbrt left-side-a 66 middle: left-side-a: word-initial: left-side-b: cat pref left-side-a 34 cat root left-side-a 133 cat derv right-side-b 67 cat hyph word-final 1024 cat infl word-final 1056 cat lbrt middle 1155 cat pref left-side-b 34 cat hyph word-final 1024 cat pref left-side-b 34 cat derv right-side-a 1027 cat lbrt middle 2083 cat root left-side-a 1093 cat hyph word-final 1024 cat infl word-final 1056 The MITalk Decomp program performed its task quite well; it could analyze 95% of running text \[Allen (personal communication) \]. In order to achieve this level of performance, the authors of Decomp made a conscious decision not to deal with stress alternations (festive I festivity), vowel shift and tensing (divine / divinity), and other phonological rules associated with latinate morphology. Basically, there was only one rule for combining the pronunciations of morphological pieces: simple concatenation with a few simple rules to account for spelling alternations at the juncture:  All affixes were assumed to be stress neutral. Words like festivity and divinity which require a richer understanding of the interaction of morphology and phonology were entered into the lexicon as exceptions.</Paragraph>
      <Paragraph position="1"> The decision not to handle more complicated morphological and phonological rules was based on the belief that it is hard to do an adequate job and that it wasn't necessary to do so because the rules are not very productive and hence it is possible (and practical) to list all of the derived forms in the lexicon. I'd like to believe that morphology and phonology have progressed enough over the past ten years that this argument does not have as much force as it did. Nevertheless, I have to admit that the payoff may be marginal, especially if measured in short term savings in the size of the lexicon and memory costs. The real value in the enterprise is more long term; I am betting that pushing the theoretical linguistic understanding with a demanding application such as speech synthesis will uncover some new insights.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="156" end_page="157" type="metho">
    <SectionTitle>
3. Types of Morphological Combination
</SectionTitle>
    <Paragraph position="0"> It has long been recognized that &amp;quot;stress-shifting&amp;quot; morphology (e.g., divin+ity) differs in quite a number of respects from &amp;quot;stress neutral&amp;quot; morphology (e.g., divine#ness). It is a well-established convention to mark the &amp;quot;stress-shifting&amp;quot; morpheme boundary with a &amp;quot;+&amp;quot; symbol and to mark the &amp;quot;stress-neutral&amp;quot; boundary with a &amp;quot;#&amp;quot; symbol. (Scare quotes are placed around &amp;quot;stress-shifting&amp;quot; and &amp;quot;stress-neutral&amp;quot; because these terms are probably not quite right.) This paper will also use the terms Level 1 and Level 2 to refer to the two types of morphological combination, respectively. This terminology is taken from the literature on Level Ordered Morphology and Phonology (e.g., \[Mohanan\]) which argues that &amp;quot;+&amp;quot; boundary (level 1) morphology is ordered before &amp;quot;#&amp;quot; boundary (level 2) morphology and that this ordering dependency has important theoretical implications.</Paragraph>
    <Paragraph position="1"> It is worthwhile to review some of the well-known differences between &amp;quot;+&amp;quot; boundaries and &amp;quot;#&amp;quot; boundaries. Informally &amp;quot;+&amp;quot; morphemes such as in +, ad +, ab +, +al, +ity are (generally) derived from Latin whereas &amp;quot;#&amp;quot; morphemes such as #ness, #1y come from Greek and German. This historical trend is only a rough correlation and has numerious counter-examples (e.g., the German suffix -ist behaves like &amp;quot;'+&amp;quot;). The program uses the following set of prefixes and suffixes: * Level 1 &amp;quot;+&amp;quot; Prefixes: a, ab, ac, ad, af, ag, al, am, an, ap, at, as, at, bi, col, corn, con, cor, de, dif, dis, e, ec, ef, eg, el, em, en, er, es, ex, ira, in, ir, is, ob, oc, of, per, pre, pro, re, suf, sup, sur, sus, trans * Level 1 &amp;quot;+&amp;quot; Suffixes: ability, able, aceous, acious, acity, acy, age, al, ality, ament, an, ance, ancy, ant, ar, arity, ary, ate, ation, ational, ative, ator, atorial, atory, ature, bile, bility, ble, bly, e, ea, ean, ear, edge, ee, ence, ency, ent, ential, eous, ia, iac, ial, ian, iance, iant, iary, iate, iative, ibility, ible, ic, ical, ican, icate, ication, icative, icatory, ician, icity, icize, ide, ident, ience, iency, ient, ificate, ification, ificative, if y, ion, ional, ionary, ious, isation, ish, ist, istic, itarian, ite, ity, ium, ival, ive, ivity, ization, ize, le, ment, mental, mentary, on, or, ory, osity, ous, ular, ularity, ure, ute, utive, y * Level 2 &amp;quot;#&amp;quot; Prefixes: anti, co, de, for, mal, non, pre, sub, supra, tri, ultra, un  * Level 2 &amp;quot;#&amp;quot; Suffixes: able, bee, berry, blast, bodies, body, copy, culture, fish, ful, fulling, head, herd, hood, ism, ist, ire, land, less, line, ly, man, ment, mental, mentarian, most, ness, phile, phyte, ship, shire, some, tree, type, ward, way, wise There is also a well-known precedence relation between + and #. With very few exceptions, # morphemes nest outside of + morphemes. Thus, we have non # \[in + moral\] but not *in + \[non # moral\]. The precedence relation yields some subtle (but Jcorrect) predictions. Observe that -able can be a level 1 affix in some cases (e.g., cbmparable) and a level 2 affix in others (e.g., emplbyable). Notice the contrast between INcomparable and .UNexmployable; the + marked comparable takes the + marked prefix in + whereas, in contrast, the # marked employable takes the # marked prefix un#. This same contrast is brought out by the famous pair: indivisible I undividable. (This argument is no longer considered to be as convincining as it once was because of so-called bracketting paradoxes which will be discussed shortly.) Word formation rules are also sensitive to the difference between + and #. Note that + morphemes can attach to bound morphemes (e.g., crimin + al), but # morphemes cannot (e.g., *crimin #ness, *crimin # ly, *crimin # hood). In addition, # morphemes attach more productively than + morphemes.</Paragraph>
    <Paragraph position="2"> &amp;quot;It is clear that #ness attaches more productively to bases of the form Xous than does +ity: fabulousness is much &amp;quot;better&amp;quot; than fabulosity, and similarly for other pairs (dubiousness I dubiety, dubiosity). There are even cases where the +ity derivative is not merely worse, but impossible acrimonious I *acrinoniosity, euphonious I *euphonosity, famous I *famosity. There is also the simple list test, which is still a good indicator. Walker (1936) lists fewer +ity derivatives than #ness derivatives of words of the form Xous.&amp;quot; \[Aronoff, pp. 37-38\].</Paragraph>
    <Paragraph position="3"> Aronoff continues to point out that the semantics of # boundaries tend to be more predictable and compositional than + boundaries. The meaning of callousness, for example, is more predictable from the meanings of callous and ness than the meanings of variety, notoriety and curiosity are from the meanings of their parts.</Paragraph>
    <Paragraph position="4"> The following list summarizes some of the differences between +  The remainder of the paper will be divided into two sections, the first will be concerned with level 1 morphology and the second with level 2 morphology and compounding. Level 1 morphology has been studied more heavily in the lingusitics literature; level 2 is perhaps more important for practical applications, at least in the short term.</Paragraph>
  </Section>
class="xml-element"></Paper>