Nonce words are widely used in linguistic research to evaluate areas such as the acquisition of vowel harmony and consonant voicing, naturalness judgment of loanwords, and children's acquisition of morphemes. Researchers usually create lists of nonce words intuitively by considering the phonotactic features of the target languages. In this study, a corpus of Turkish orthographic representations is used to propose a measure for the nonce word appropriateness for linearly concatenative languages. The conditional probabilities of orthographic co-occurrences and pairwise vowel collocations within the same word boundaries are used to evaluate a list of nonce words in terms of whether they would be rejected, moderately accepted or fully accepted. A group of 50 Turkish native speakers were asked to evaluate the same list of nonce words. Both the method and the participants displayed similar results.
Nonce words are frequently employed in linguistic studies to evaluate areas such as
well-formedness [
The acceptability of nonce words can be investigated by experimental investigations
through phonotactic properties or factor-based analysis [
In this study, the acceptability of nonce words was assessed using the conditional
probabilities of the bigram co-occurrences of the orthographic representations locally
and the pairwise collocations of the vowels within the same word boundaries. Similar
methods within the context of phonotactic modeling had been used for Finnish vowel
harmony [
Turkish has 8 vowels and 21 consonants, and it is agglutinative with a considerably
complex morphology [
(1) Ko¨pek adam-ı ısırdı.
Dog man-Acc bit The dog bit the man.
Ko¨pe˘g-i adam ısırdı.
Dog-ACC man bit
The man bit the dog.
The description of Turkish word structure depends heavily on morphophonological constraints and morphotactics. In Turkish morphotactics, the continuation of a morpheme is determined by the preceding morpheme or by the stem as in (2). (2) ev-de-ki house-Loc-Rel The one in the house
*ev-ki-de
These morphotactic constraints in Turkish are captured by statistical models based
on conditional probabilities [
Vowel harmony is dominantly effective in Turkish morphophonology in order to preserve the roundedness and the frontness of vowels within the same word boundaries. While a morpheme with a vowel is concatenated to a string, its vowel is modified with respect to the roundedness and frontness properties of the most recent vowel in the string as in (3).
(3) ev-ler house - Plu houses oda-lar room - Plu rooms bil-di know - Past knew duy-du hear - Past heard
Another important phenomenon in Turkish morphophonology is voicing. If some of the strings terminating with the voiceless consonant, ‘p, t, k, ¸c’, are followed by the suffixes starting with vowels, then the consonants are voiced as ‘b, d, ˘g, c’ as in (4). (4) sonu¸c result sonuc-um result -1S.Poss my result
Consonant assimilation is also important in Turkish morphophonology. The initial consonants of some morphemes undergo an assimilation operation if they are attached to the strings terminating in the voiceless consonants, ‘p, t, k, ¸c, f, s, ¸s, h, g’, as in the surface forms of the Turkish past tense -DI in (5).
The final Turkish morphophonological phenomena that need to be briefly mentioned are deletion and epenthesis occurring as in (6).
(6) hak right hakk-ım right - 1S.Poss my right isim name ism-im name - 1S.Poss my name
The Turkish morphophonological phenomena described above occur in the cooccurrences of the orthographic representations in the concatenating positions except in vowel harmony and the deletion. This results in high conditional probabilities evaluated using the frequencies of the pairs of consecutive orthographic representations. Since the vowel harmony and deletion take place after or before the concatenation positions, their pairwise collocations within the same word boundaries are also required to be utilized in the statistical model.
The transition probability between A and B is simply based on the conditional probability statistics as in (7). (7) P (B|A) = (frequency of AB) / (frequency of A)
Infants are reported to successfully discriminate speech segments using transitional
probabilities of syllable pairs [
Statistical approaches to linguistics support the empiricist view; and they provide an explanatory account of linguistic phenomena such as the decrease in performance errors and language variations. Considering the properties of the Turkish language, using the conditional probabilities of orthographic representations and the collocations of vowels within the same word boundaries is a plausible method to decide whether nonce words or loan words will be rejected, moderately accepted or accepted 3
Let s be a string such that s = u1u2. . . un, where ui is a letter in the Turkish alphabet. The string s is unified with the empty strings σ and ε such that s = σu1u2. . . unε, where σ denotes the initial word boundary and ε denotes the terminal word boundary. The overall transition probability of the string s is evaluated from the METU-Turkish Corpus using Formula 1.
n+1 Pt(s) = Y P (ui|ui−1) (1)
1
For example, using the Formula 1, P (a|σ) gives the probability of the strings starting with the letter a, and P (b|a) estimates the probability of the substring ab in the corpus. Now let v be a subset of the string s such that v = ui,1uj,2 . . . uk,m where uk,m is the mth vowel in the kth location of the string s. The overall vowel collocations of the string s are estimated from the substring of vowels v using Formula 2.
Pc(v) = Ym g(vi−1vi)
f (vi−1) 2
if |v| > 1 Pc(v) = f (vi) if |v| = 1 (2)
In the Formula 2, the function f (vi) gives the frequency of the words that contain the vowel vi as a substring in the corpus. The function g(vi−1vi) gives the frequency of words in which the vowels vi−1 and vi are collocating not necessarily in immediately consecutive positions but within the same word boundaries. The acceptability probability of the string s is calculated by Pa(s) = Pt(s)Pc(v). The acceptability decision of the string s in the method is made by using the Formula 3.
Pa(s) ≥ 10−(t+v) 10−(t+v+1) > Pa(s)
10−(t+v+1) ≤ Pa(s) < 10−(t+v) (3) where t is the number of transitions (which is the length of the string + 1) and v is the number of the vowel collocations (which is the number of the vowels - 1) in the string. If the string s has only one vowel, then v = 1.
The method was applied to the list of nonce words given in the following section. The same list was also given to the 50 Turkish native speakers to evaluate the acceptability of each item. The comparison of the results from the method and the native speakers is given below. 4
The nonce word talar is evaluated as in (8) (8)
Pa(talar) = Pt(σtalarε)xPc(aa) = P (t|σ)P (a|t)P (l|a)P (a|l)P (r|a)P (ε|r)xPc(aa) = 7.66e − 06xPc(aa) = 7.66e − 06 ∗ 4.75e − 01 = 3.63e − 06
Since Pa(talar) ≥ 10−(6+1), in which 6 conditional probability estimations and 1 vowel collocation are evaluated, the nonce word talar is accepted. The word list was evaluated by the 50 selected Turkish speakers. The distribution of the native speaker responses and the results of the method are given in Table 1.
For 82% of the words the Turkish native speaker’s responses are in agreement with the results from the method. The method failed to simulate the responses from the participants in 18% of the results. 5
The acceptability of loan words and nonce words is mainly determined by the
phonological properties of the target language and the current approaches are
syllablebased [
The nonce word u¨lu¨ was rejected by the method but accepted by the participants. A possible reason might be that the nonce word u¨lu¨ sounds similar to an existing Turkish word o¨lu¨ ’death’. Similarly, the responses for the nonce word nort were in disagreement. This nonce word has a similar pronunciation to an English word north and the most of the participants also knew English as a foreign language. Therefore, the participants might also make use of their foreign language knowledge to evaluate nonce words.
Although the method does not assume to utilize any property of Turkish phonology and it does not implement any phonologic filtering mechanism, it is able to mimic, in a remarkable way, a large number of the responses from the participants. Indeed, this study does not propose that acceptability is based on raw orthographical representations rather than syllables and phonemes. Instead, it underlines that simple pairwise conditional properties and vowel collocations from a corpus can give an estimation of the acceptability of a list of nonce words. This can be used by researchers that need an evaluation for the nonce words for their studies when no phonologically annotated corpus with syllables exists. 6
The method needs to be tested with larger word lists. The method is successful because there is a close correspondence between phonotactics and orthotactics in Turkish. It requires improvements in terms of the morphophonological properties of target languages. The method uses exact orthographic representations. Thus, it requires an additional phonological similarity measure for the representations to increase the success rate.
The threshold values for the acceptability decisions depend on word lengths. They also need to be improved with respect to the target languages. The method also needs to be tested and adapted for the languages with ablaut or umlaut phenomena such as English and German, and the templatic languages such as Arabic and Hebrew.