C. Anton Rytting and Dr. Royal Skousen, English; Deryle Lonsdale, Linguistics
The k/Ø Alternation and Analogical Modeling
Turkish is an agglutinative language, meaning that each word consists of a word stem followed by a variable number of suffixes. Sometimes, the addition of suffixes causes the stem to change its form. For example, stem-final velar consonants (i.e., /k/ and /g/) tend to be deleted between a stem vowel and a suffix vowel:
This conditional deletion of stem-final velars is known in the literature as the k/Ø alternation (Zimmer & Abbott 1978, Sezer, 1981) or as velar drop (Inkelas & Orgun 1995).1 However, there are some known exceptional cases to the k/Ø alternation, where velar deletion does not apply. The majority of these cases fall into two categories (see Sezer 1981).
Sezer (1981) postulates a rule that deletes final /k/ after a vowel in polysyllabic words. Inkelas and Orgun (1995) provide theoretical justification for exception #1, by linking it to a previously proposed language universal (i.e., a condition thought to apply to all languages). In addition, Zimmer and Abbott (1978) provide empirical evidence that exception #1 is productive, by showing that native speakers apply it to nonce (or made-up) words. However, neither Inkelas and Orgun’s analysis nor Zimmer and Abbott’s psycholinguistic experiment explicitly or completely account for Sezer’s second class of exceptions: loan-words with long final vowels.
Moreover, as rules have not always proved adequate to describe the complexities of human language use, several attempts have been made to find non-rule explanations for linguistic phenomena. One such approach is Analogical Modeling of Language (AML), developed by Dr. Royal Skousen.2 This approach uses the characteristics and behavior of common words, which are already “learned” and accessible in the speaker’s memory, to predict the behavior of less common words. However, Skousen’s theory has yet to be tested on a wide variety of problems.
Methodology
In order to test the predictive power of Skousen’s AML algorithm, I created a data set of 1440 word stems which ended in /k/ or /g/. These word stems were all listed as “actively used” in the “Turkish Electronic Living Lexicon” or TELL.3 All the entries in TELL were based on the responses of a 63-year old native speaker of a standard Istanbul dialect.
The 71 velar-final word stems which were listed as “not actively used” in TELL were used as test items, to see if Skousen’s AML algorithm would predict the form given in TELL. Within this word list were sufficient examples to test both the “normal” case and exception class #1. In addition, two other word lists were tested: a list of Arabic loan words with long final vowels, taken from Sezer (1981) to represent exception class #2, and a set of nonce words used by Zimmer and Abbott (1978). These words were also tested, to see if the AML algorithm would accurately predict Sezer’s claims and the responses given by Zimmer and Abbott’s subjects.
Results
In the first test, all of the words which deleted the final velar (the “normal” case) and all of the monosyllabic words (exceptional class #1) were predicted correctly. These two classes of words were also correctly predicted by the universal cited by Inkelas and Orgun (1996). Thus, the AML algorithm is equal to the rule-based approach in its predictive power for these words. For the second exceptional class, all of Sezer’s words were correctly predicted when the final vowel was treated as long. AML also confirms Sezer’s claim that vowel length is crucial: “uncommon” words failed to be predicted correctly when the final vowel was treated as short.
In the third and final test, the nonce experiment, twenty-one of the twenty-four words were predicted “correctly” (i.e., as the majority of the subjects had them), and three were missed. This is not remarkably unusual, for about half of the subjects also “missed” three or more of the twenty-four items. The deviance seems explainable from individual characteristics of the TELL database’s informant, whose speech differs from the “average” speaker (see below).
An unanticipated class of exceptions
There is a final category of exceptional words which has not yet been addressed. About 11% of the polysyllabic words with a short final vowel in the TELL lexicon retained the final velar where both Sezer’s rule and the implications of Zimmer and Abbott’s experiment would predict velar deletion. These “exceptional” words were also mis-predicted by the AML algorithm. This is still an unresolved issue. It may be just an idiosyncrasy of the TELL informant. But if they prove more wide-spread, then these “unexplained” exceptions show that both rule-based approaches and this application of Skousen’s Analogical Modeling have room for further refinement. Nevertheless, this study has shown that Analogical Modeling of Language is equal in predictive power to the rule approaches proposed so far.
References
- Zimmer, Karl E. and Barbara Abbott. 1978. The k/Ø Alternation in Turkish: some experimental evidence for its productivity. Journal of Psycholinguistic Research 7.35-46. Sezer, Engin. 1981. The k/Ø alternation in Turkish. Harvard studies in Phonology, ed. By George N. Clements, 354-82. Bloomington: Indiana University Linguistics Club. Inkelas, Sharon and Cemil Orhan Orgun. 1995. Level ordering and economy in the lexical phonology of Turkish. Language 71:4.763-793.
- Skousen, R. 1989. Analogical modeling of language. Dordrecht, The Netherlands: Kluwer.
- For more information, see http://socrates.berkeley.edu:7037/AboutTELL.html