<?xml version="1.0" standalone="yes"?>
<Paper uid="J79-1019">
  <Title>fi Case History in Computer Exploration of Fast Speech Rules</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
A CASE HISTORY
GDMPUTER EXPLORATION
FAST SPEECH RULES
</SectionTitle>
    <Paragraph position="0"> In conversational speech, words run toq6t her and interact causing thei phonological f~rms to differ from their citation forms. Fast speech rules attempt t3 qescribe these changes as speech becomes faster and more casual, In developing 3 ny set of phonological rules, computerized grammar testers are a useful and important aid. They necessitate (3. precise, aonsistent Eornulation of the rules and allow the generation of s-ple d.erivations. In applying these ~ules to a diverse set of utterances, we can first confirm that the rules really dc apply where we expect them to, and then experiment vith various rule orderings to obser~e their effects.</Paragraph>
    <Paragraph position="1"> The Phonological Grammar Tester of Friedman and Morin was use-d to test two sets Of* fast spe'ech rules from the ARQA Spsech Understanding Research ccmmunity. Working with C hese rules lad to certain observations about the interactions and nature of fast speech rules in general. 3x1 addition to testing these two sets of fast speech rules, we were also interested in the problems of teSting such a grammar vith this program.</Paragraph>
    <Paragraph position="2"> Appehdices incltlde an overview of the grammar tester and finaJ output from our testing,  In developing any sebt of phonological rules, computerized grammar testers are a useful and important aid. They necessitate a precise, consistent formulation of the rules and aL1ow the generaticn of sample derivations. In applying th~se rules to a diverse set of utterances, we can first confirm that the rules really do 3pply where ue expect them to, and then experiment with various rule orderings to observe their effects, Tradit iona 1 phonological rules describe transitions from the underlying form to the surface form of a word. However, in casual (fast) speech, words are not distinct, independent units, but run together and interact. Fast speech rules atte~pt *-o describe these changes that occur as speech becomes faster and more casual (4, 8, 9).</Paragraph>
    <Paragraph position="3"> This is a. case #history of the examination of two particular sets of rules on the Phonological Grammar Tester (PGT) of Friedman and Borin (3) . An overview is given in Appendix A and a sam~le of the output in Appendix 8. In addition to testing these two sets of fast speech rules, we were also interested in the problems af testing such a grammar with this program.</Paragraph>
    <Paragraph position="4"> Testing the Fast Seeech Rules -- --I- ---CmIIILI.I.</Paragraph>
    <Paragraph position="5"> For our f j cst test, w chose the fast speech rules of Neu (6) as the best available set of rules given in a reasonably.</Paragraph>
    <Paragraph position="6"> consistent notation. The rules Mere given in two formats: one usidg phonemes and the other using distinctive features. The first task was to adjust the notation into a machine usable form. For the purposes of this test, we used the feature systeh of Chomsky and Halle ('I), with three types of boundaries: syllable (+ , morpheme (3) and word (#I) . For example, the f eatur'es anterior -----a and ------ coronal had to be substituted for &amp;ace' _of art i~u1ati.o~. Also, the rules had to be modified so that each featur-e had the same type of value (integer, binary, gr markedunmarked) th~ongbout the grammar. For example, in the PG?, lstress is not equal to +ssess and is less than =stress, which ----- f could cause manf unexpected problems in the grammar. Another minor problem was aetermining whether a C I in the rules meant not ----.-- a vo we1 , I+ consl, or up-- vocall.</Paragraph>
    <Paragraph position="7"> Sample derivations were given with the rules, and these. were .used as a first test. We quickly found a number of small problems, most being that the rule had been incorrectly or incompletely stated. For example, Geminate Reduction and Alveolar Flapping would not apply across the word boundary in &amp;quot;want to&amp;quot;, while the E in &amp;quot;captain1v blocked the application of lott tali eat ion.</Paragraph>
    <Paragraph position="8"> More sample words were taken from. Neu and several more small erro-1s were faund, such aa Nasal Consonant Deletion bleeding Transitional Stop Insertion and thereh y indi'cating that it should be ordered after Transtop. We still had SOB-e unresbl ved problems, such as &amp;quot;Ts /f e d a 1/ a proper fast speech form of 'f.ederal'?lg, when the revisions and additions (7) arrived, Many of thes e revisions corresponded to our solutions of problems encoun\telCed during testf ng of these rules. The f iaal form of this first sot is given in lppendix C.</Paragraph>
    <Paragraph position="9"> Since there was little obvious order in the rules, graphing the ~artial ordering given was very helpful in exploring the rule ordering. The partial ordering given with the rulas, Ficrure 1, was quickly refined to that in Figure 2 by testing the sample der i va't i cns .</Paragraph>
    <Paragraph position="10"> In these new 'rules, particular attention was paid to the presence and posit ion of syllable, aorphene and word boundaries in the environments for which the rules should apply. Other fast speech rules from members. of the ARFA Speech Understandipg Research (SUR) community (11, 12, 13) vere used 3s references, especially in respect tb boundary position, in impledenting fh ase new rules. In testing the rules, most of the problems involv'ed the positioning and poesence of boundari-es.</Paragraph>
    <Paragraph position="11"> The second Dental Deletion rule was dropped from the grammar because it appeared tc duplicate the effect of assimilating an alveolar flapped g to the preceding 5.</Paragraph>
    <Paragraph position="12"> We had few transcriptions of fast speech forms, so in</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
GRAPH OF PARTIAL ORDERING
</SectionTitle>
    <Paragraph position="0"> testing these rules a major problem was in determining wha-t did and did not constitute proper fast speech forms. In slow speech, it is fairly easy to break up the utterance into a sequence of phonemes, but in fast speech, the utterance is more contiquous and it is harder to decide what is a segment and wh.ere the segments begin and end. For example, the palatalization rules produce the following derivation:</Paragraph>
    <Paragraph position="2"> We are not inclined to believe that an intermediate form /did##Yyu/ exists between steps (a) and (b) in this derivation.</Paragraph>
    <Paragraph position="3"> I think that (b) is a true intermediate form, existing in slower, casual speech, but not in fast speech.</Paragraph>
    <Paragraph position="4"> Our best evidence indicated that the RUH is the same as the syllabic &amp;quot;rU I thus raising the problem of ordering the Syllabicizing and Ilu h-rsduct ion rules. Are the proper forms /g u v +- n r/ ahd /f e d + r l/ or, more likely, /g u v + n r/  and /f e d + r I/? Qc, could it possibly be /g u v + n r/ and /f e d + r I./, with the Syllabicizing rule not deleting the schua, but naking it very short and passibly devoicing it? Another disturbing result of the testing vas the application of rules across deleted segments. For exaaple, after qqvernqg has been reduced to /g u v + n r/, Proqressive Assimilation applies, producing /g u v + m r/. Similarly, ---- sanitq ------- clause becomes /s ae n + t a .../ which becomes /s ae n + f' a .../, closing in on a .fast speech form of santa Claus. The final form of the second set of rules and the output are given in Appendix D 6 Suggested &amp;dific_gtions to the PGT - -----.. - ---------I-The PGT was writfen for the testing of traditional ph~nologicat grammars which take an underlying form to a Burface form, and thus fast speech rules present several unforeseen problems.</Paragraph>
    <Paragraph position="5"> Fast speech r'ules do not constit.ute a conpleke grammar, but are iostead interspersed, among the traditional rules, Ln producing an intermediate or surfhce for^ for input to the grammar, the use of diacritics would praduce 2 simpler and more econcmical represent a tion st the input For example, in the present system, it is necessary to give a separate definition for each stress assignment for each vowel, instead of defining each towel cnce and assigning the stress through the use of diacritics.</Paragraph>
    <Paragraph position="6"> Syllabification aDpears to occur before some of the rules, an d after others. TO avoid having to insert, rules for syllabificatio~~ or having the rules preceding syllabification ignore syllable boundaries in the input, the PGT should have a l@human intervention&amp;quot; rule which would allow the user to chmge the tree as part of the derivation, simulating the effect of the WmissingfI rules. &amp;quot;Human interventionn rules could be used to postFone Qriting rules, either ro a later session or indefinitely if the rules are too hatd to write or if they are &amp;atsi.de the scope of the problem.</Paragraph>
    <Paragraph position="7"> Another problem arises from most, if not all, fast spzech rules being optional. The present system generates a single derivation using a ando om number generator to decide whether or not- to apply optional rules. As a result, to obtain a good best of the effect of an optional rule in tho grammar, we have to do two tests, Ohe with the rule obligatory and one without the rule. With. a large number of optional rules operating on a rron-trivial set of test data, having to test each possible rule combination individually would be a- staggering task. A more satisfactory method wquld be for the PGT to produce a set of derivations such that any tine an optional rule can apply, the PGT produces a derivation for the case where the. rule was applied and one for the case where it wasn't. Cohen ana Mercer (2) have implemented such a feature in their rule tester by storing the result of the application of a rule as a directed gra.ph. Subsequent rules are applied to all paths through the graph, producing a neu graph. In considering the problem of when an optional rule should be applied, we are considering pgfouance (10) . To this end, we might be better served by having the variable rules of Iabov (5) ia addition to traditional phonologica1 rules. f nstead of the all- or- nothing matching of normal rulas, variable rules employ incremental matchins, with the likelihood of the rule being applied being a function of the degree of match, the rate of speech, and the nature of the conversation (e,g. Is it formal or casual?) . For example, for: the Froqrelssive Assimilation mile for point of articuigtion (PROARTICJ , /s O f n/ ((soften) becomes /S (3 f m/ only in very fast speech, hut /p r + f y u + n/ v N (profusion) becomes /p c + f y u + z n/ in sost fast speech. In additian, for each variable rule applied, the PGT should use the associated probabili tv tc produce a Einal probability for each, derivation. Unfortunatelyr the ilbpleaentation of variable rules would necessitate major changes to the PGT.</Paragraph>
    <Paragraph position="8"> Convenient, but not necessary, would be the ability to make a rule &amp;quot;blindw to certain types of nodes, For example, Chomskp and Halle (1) allow sosle boundary symbols to be invisible to their phonological rules. Tn some cases, this ability might be able to fulfill the function of a &amp;quot;human interventionv rule, Observat ions about Fast Skeech R ules -------------.---- ------Testing the fast speech rules led ta certain insights into the interactions of the rules. The most vexing problems verbe with the representation of boundaries. The use of 2, &amp;, and gt seems artificial and cumhersome, A moce natural approach seems to be to assign the boundaries varying degrees of $treggf_h, and then allow the phonolcgical rules to operate on the strength assignments. Consider the problem of HJohngs going to uorkll becoainu the fast speech form l'Johnts qonna work1' in the case where qoinq is an auxiliary but not in the case uhere it is the verb. Rather than having the phonological rule sensitive to the syntax, -it seems much cleaner to have it sensitive tp the boundary strength between on and to. To this end, we must allow the syntax to influence the initial assignaents of boundacy strength. The use of boundary strengths also seems to solve the problem, of how boundaries are reduced. For example, in reducing want to to wanna should the word boundary be reduced to a me--- ---I morpheme or a syllable bounda'ry? Another ase of boundary strengths might be to mark the position of segments deleted fcom a word by strengthening the neighboring boundaty. For example, when governor ------- is reduced to sv I ner, the p is lengthened, indicating the possible presence of a stronger then normal syllable boundary. This stronger boundary. would then prevent the from changihg the g to an - m under Progressive Assimilation. An alternative is to hare a &amp;quot;strong syllablew boundary symbol which would be &amp;quot;strongerwq than a &amp;quot;weak wordm bounaary symbol, but this woulil be mdssy and unnatural.</Paragraph>
    <Paragraph position="9"> It voola probably be more productive to represent fast speech forms with phonemes utilizing multi-valued (1 through 7 instead of + and ) features. Giving the segments length attributes would allow segments to be deleted by a combination of rules which reduce segment lengths. Also, in conparing forms, a very short seqment could be considered to be deleted.</Paragraph>
    <Paragraph position="10"> Fast speech rules seem to be selective as to the words to which they can apply. Zwicky (14) gives examples and counter-examples for various fast speech rules; However, it seems that the counter-examples are words which do not ftequently occ,ur in &amp;aqua1 speech, while the examples were common. This ooint was illustrated when , in an early run, that Schva Deletion produced /k ae n + g r u/ from 'tkangaroow, AT first, this East speech form was unacceptable, but with use, it becdme more and more acceptable. Interestingly enough, a small random polling se;ned to indicate that this form is acceptable &amp;quot;if yout re an AUS traliantr . That fast speech rules apply t~ what 1s common in the speaker's vocabulary is also evidenced by the heavy application of fast speech rules to local names.</Paragraph>
    <Paragraph position="11"> The similarity af the palatization rules in this fast spaecfi grammar and in normal Fnglish (1) raises tha question of What are 'fast speech rules?&amp;quot;. If normal rules are restricted versions of the more general processes which also account for the Past speech forms (14), then we must be careful in using slow speech surface forms to test the fast speech rules. For example, part of the data covered by the Schwa Deletion rule might actuslly result from schwas being *ser&amp;gq between consonants in slow speech formS.</Paragraph>
    <Paragraph position="12"> york was supported undec National Science Fou\naaaiion 1309, with the supervision and quidanc~ of Joyce</Paragraph>
  </Section>
class="xml-element"></Paper>