USING AN ON-LINE DICTIONARY TO FIND RHYMING

consonant clusters which, especially when they containnasals or liquids, result in less successful rhymes. A thirdstudy will allow us to identify "discontinuity" in therhyme List, across which rhyming words ate very unlikelyto be found. In Figure 1., a discontinuity seems to occurbetween currency and trustworthy.The second encoding step arranges the segments for apronunciation in the order representing theft importancefor determining rhyme. This ordering is also the subjectof continuing experimentation. The current arrangement is as follows:The second reason that our rhyme lists ate not perfectis that it is unlikely that any single dimension will besufficient to guarantee that all and only good rhymes fora given word will appear adjacent to that word in thedimension's order, if only because different people disagree on what constitutes "good" rhyme.(1) All segments preceding the syllable bearing primary stress are recorded in the order that they occurin the pronunciation string.(2) All consonantal segments in and following thesyllable beating primary stress are added to the encoding in the order in which they occur.Examples(3) All vocalic segments (vowels and diphthongs) inand following the syllable bearing primary stress areplaced before any segments for trailing consonantsin the final syllable. [f there are no trailing consonants in the final syllable, then these vocalic segments are placed at the end of the encoding.We give two sequences of words selected from theWordSmith RHYME onarymissionaryNote that this scheme preserves the order of the segments preceding the point of primary stress, as well asthose in the final syllable. For words where primarystress occurs before the final syllable, the vowels areraised in importance (with respect to rhyming) over allconsonants except final ones. This procedure allows usto capture the intuition that fervency is a better rhymefor urgency than agency.These fi, words have their primary stress in the forthsyllable from the right, and they also have the same fourvowel sounds from that point onwards. Notice that thespelling of antiphonary would place it quite far from theothers in a standard reverse dictionary. In addition, theextra syllables at the beginning of antiphonary,seditionary, and expeditwnary are irrelevant for determining rhyme.The final step in the encoding procedure reverses thephonetic segment strings right-for-left, groups them according to the position of the syllable bearing primarystress (i.e., the distance of that syllable from the end ofthe word) and sorts the groups just as in the productionof reverse dictionaries. The difference is that nowneighbors in the resulting sorted list have a better chanceof rhyming because of the use of pronunciations and theapplication of our intuitions about rhymes.writewrightriterightThese four words, each a homonym of the others, sharea single record in the rhyming index and are thereforeadjacent in the WordSmith RHYME dimension.We note that the resulting lists of rhymes are not perfect.This is so first because we have not completed the experiments which will result in an "optimal" set of intuitions about the encoding process.One plannedexperiment will clarify the position of the schwa vowelin the vowel triangle. Another will study intervocalic3. Pronunciation of Unknown WordsReading aloud is a complex psycholinguistic process inwhich letter strings ate mapped onto phonetic repres 280

entations which, in turn, are converted into articulatorymovements. Psycholinguists have generally assumedThere are of c o u r s e many unanswered questions abouthow readers actually generate pronunciations by analogy. One approach to answering the questions is tobuild a computational system that can use various strategies for finding lexical neighbors, combining partial(Forster and Chambers, 1973) that the mapping fromletters to phonemes is mediated by two processes, onebased on rules and the other based on retrieval of storedpronunciations. For example, the rule ea - / i / converts the ea into the long e sound of leaf. The otherprocess, looking up the stored pronunciation of a word,pronunciations, etc., and then compare the output of thesystem to the pronunciations produced by human readers. The following is an outline of such a computationalsystem.is responsible for the reader's rendering of deaf a s / d f / ,despite the existence of the ea - / i / r u l e .Both processes are believed to operate in the pronunciation ofknown words (Rosson, 1985).Two WordSmith files will be used to support a proposedprogram that generates pronunciations for unknownwords based on stored pronunciations of known words.The fh'stis a main filewhich is keyed on the spelling ofwords and which contains pronunciations organized according to part of speech. This is the file which supported the P R O N U N Cand R H Y M EWordSmithUntil recently, it was generally assumed that novelwords or pseudowords (letter strings which are not realwords of English but which conform to English spellingpatterns, e.g. heat') are pronounced solely by means ofthe rule process because such strings do not have storedfunctions described earlier. In this file,each pronunciation of a word has stored with it a mapping from itsphonetic segments onto the lettersof the spelling of therepresentations in the mental lexicon.Hcwever,Glushko (1979) has demonstrated that the pronunciation of a pseudoword is influenced by the existence ofword. These mappings were generated by a P R O L O Glexical "neighbors." i.e., real words that strongly resemble the pseudoword. Pseudowords such as heal, whoseprogram that uses 148 spelling-to-pronunciation rulesfor English (e.g.ph - /f/). The second fileis an indexto the main file keyed on reverse spelling. This file isequivalent to the one which supports the R E V E R S EWordSmith dimension shown in Figure I.closest neighbors (leaf and deaf) have quite differentpronunciations, take longer to read than pseudowordssuch as hean, all of whose close neighbors have similarpronunciations (dean, lean, mean, etc.). (It has beenassumed that words which differ only in initial consonants are "closer" neighbors than those which differ inother segments.) Giushko has also demonstrated an effect of lexical neighbors on the pronunciation of familiarwords of English.The strategy for generating a pronunciation for an unknown word is to find its lexical neighbors and producea pronunciation "by analogy" to their pronunciations.The procedure is as follows: (a) Segment the spelling ofthe unknown word into substrings. (b) Match each substring to part of the spelling of a known word (orThe picture that emerges from this psychological workdepicts the retrieval process as selecting all stored wordswhich are similar to a given input. If the input is notfound in this set (i.e., the input is a novel word orpseudoword), its pronunciation is generated by analogyfrom the pronunciations that are found. Analogicalprocessing must take note of the substring common tothe input and its neighbors (ean in the case of hean), useonly this part of the pronunciation, and make provisionfor pronouncing the substring which is different (h).When the pronunciations of the lexical neighbors areconsistent, the pronunciation of the pseudoword can begenerated by the reader more quickly than when thepronunciations are inconsistent.words). (c) Consult the spelling-to-pronunciation mapto find the pronunciation of the substring. (d) Combinethe pronunciations of the substrings into a pronunciationfor the unknown word.These steps are illustrated below for the unknown wordbrange.(a) Segmentationbrange - - initial substring - - - final substringStrategies for segmentation will be discussed later.281

we conducted suggests that there is considerable disagreement over the pronunciation ofbrange. About half of those we asked preferredpronunciation (i), while the others chose (ii).(b) Matchingbran is the longest initial substring in brangethat matches a word-initial substring in the dictionary. The word bran is a dictionary entry,and 20 other words begin with this string.range is the longest final substring in brange thatmatches a word-final substring in the dictionary. The match is to the word range. In thereverse spelling Fde, 22 other words end inange.In the example shown above, segmentation is driven bythe matching process, i.e. the substrings chosen are thelongest which can be matched in the main file and thereverse spelling f'fle. There are, of course, other possiblestrategies of segmentation, including division at syllableboundaries and division based on the onset-rhymestructure within the syllable (for brange, br angelEvaluation of these alternative methods must await fur-(c) Pronunciation of substringsAll 21 words that have the initial string matchfor bran have the mappingb r a nIIIIbraDnther experimentation.There are other outstanding questions related to theMatching and Combining steps. If matches cannot befound for initial and final substrings that overlap (as inthe example) or at least abut, then information about thepronunciation of an internal substring will be missing.Finding a match for an internal substring requires eithera massive indexing of the dictionary by letter position, atime consuming search of the standard indexes, or thedevelopment of a clever algorithm. With regard tocombining substring pronunciations, the problem of primary stress assignment arises when primary stress is absent from all of the substrings or is present at differentlocations in two or more of them. Finally, there is aquestion of the weight that should be assigned to alternative pronunciations generated by this procedure.Should a match to a high frequency word be preferredover a match to a low frequency word? Is word frequency more important than the number of matchingsubstrings which have the same pronunciation? Theseare empirical psycholinguistic questions, and the answers will no doubt help us generate pronunciations thatmore closely mirror those of native English speakers.In 20 of the 23 words that match word-finalange, the mapping isaInIgeI n j as in range(/renj/)The other three words are flange (/aenj/), orange ( / I n j / ) , and melange ( / a n j / ) .(d) Combining pronunciationsFrom the substring matches, the pronunciationsof/b/,/r/,/n/,/g/,a n d / e / a r e obtained ina straightforward manner, but pronunciation ofthe vowel a is not the same in the bran and angesubstrings. Thus, two different pronunciationsemerge as the most likely renderings of brange.(i) below is modelled after range or change, and(ii) is modelled after bran or branch.(i) b r a n g eI I I I Ib r e n j(ii) b r a n geI I I I Ib r n jHere, pronunciation by analogy yields twoconflicting outcomes depending upon the wordmodel selected as the lexical neighbor. If people use similar analogical strategies in reading,then we might expect comparable disagreements in pronunciation when they are asked toread unfamiliar words. A very informal survey4. C o n c l u s i o nThe two applications described here, finding rhymingwords and generating pronunciations for unknownwords, represent some ways in which the tools of computational linguistics can be used to address interestingpsycholinguistic questions about the representation ofwords.They also show how answer to thesepsycholinguistic questions can, in turn, contribute to282

work in computational linguistic.s, in this case to development of the WordSmith on-line dictionary.Rickert, W.E. (1978), Rhyme terms.35-46.AcknowledgementsStyle, 12(1),Rosson, M.B. (1985), The interaction of pronunciationrules and lexical representations in readingaloud. Memory and Cognition, in press.We are grateful to Barbara Kipfer for her preliminarywork on the syllabification of unknown words, and toYael Ravin and Mary Neff for comments on earlier versions of this report.Thomas, J., Klavans, J., Nartey, J., Pickover, C., Reich,D., and Rosson, M. (1984), WALRUS: A development system for speech synthesis. IBMResearch Report RC-10626.ReferencesWalker, J. (1924), The Rhyming Dictionary, Routledgeand Kegan Paul, London.Forster, K. and Chambers, S. (1973), Lexical access andnaming time. Joutmal of Verbal Learning andVerbal Behavior, 12, 627-635.Webster'sGiushko, R. (1979), The organization and activation oforthographic knowledge in reading aloud. Journal of Experimental Psychology, 5,674-691.Seventh Collegiate Dictionary (1967),Merriam, Springfield, Massachusetts.283

et aL, 1984), or it might be used to find rhyming words, in WordSmith's rhyming dimension, for an unknown word. Z. Rhyme The WordSmith rhyme dimension is based on two files. The first is a main file keyed on the spelling of words arranged in alphabetical order and containing the words'