Word segmentation and vowel harmony

One of the first tasks for a child in the process of language acquisition is to learn which sound sequences correspond to words in natural speech. Speech that children hear is usually composed of long strings of words and does not necessarily provide cues regarding word boundaries (Cole & Jakimik, Reference Cole, Jakimik and Cole1980). Despite that, children are universally successful at word learning, and they rarely produce word boundary errors. Thus the question how children learn which sound sequences correspond to words in a particular language naturally arises.

Previous research has shown that infants rely heavily on prosodic cues when they are segmenting words in a string of speech (Curtin, Mintz & Christiansen, Reference Curtin, Mintz and Christiansen2005; Cutler & Norris, Reference Cutler and Norris1988; Jusczyk, Reference Jusczyk1997; Reference Jusczyk1999; Jusczyk, Houston & Newsome, Reference Jusczyk, Houston and Newsome1999; Morgan, Reference Morgan1996; Thiessen & Saffran, Reference Thiessen and Saffran2003; among others). Statistical probabilities and phonotactic constraints, i.e., constraints in the order of phones, are also observed to be helpful in word segmentation (Aslin, Saffran & Newport, Reference Aslin, Saffran and Newport1998; Brent & Cartwright, Reference Brent and Cartwright1996; Saffran, Newport & Aslin, Reference Saffran, Newport and Aslin1996; among others, though see also Cairns, Shillcock, Chater & Levy, Reference Cairns, Shillcock, Chater and Levy1997; Mattys, Jusczyk, Luce & Morgan, Reference Mattys, Jusczyk, Luce and Morgan1999; Yang, Reference Yang2004). It seems natural to assume that other phonological regularities, for instance, vowel harmony, may contribute to word segmentation as well. In harmonic languages, words tend to have either all front or all back vowels. Assuming that a great majority of the words are harmonic in a language, vowel harmony may be relevant to word boundaries because shifts from one vowel type to the other (shifts from back to front or the other way around) may signal word boundaries. Indeed, studies based on Finnish show that adults can recognize words on the basis of harmony cues (Suomi, McQueen & Cutler, Reference Suomi, McQueen and Cutler1997; Vroomen, Tuomainen & Gelder, Reference Vroomen, Tuomainen and Gelder1998). In these studies, pseudo-words were created based on Finnish vowel harmony rules where word boundaries and harmony shifts overlapped and adults were observed to be sensitive to such shifts. A similar study based on Turkish vowel harmony rules was conducted on Turkish and French speakers and the role of vowel harmony vs. word stress on word segmentation was tested. The results suggested that Turkish speakers showed sensitivity to vowel sequences in word segmentation and used them together with word stress regularities, while French speakers relied only on stress cues in word recognition (Kabak, Maniwa & Kazanina, Reference Kabak, Maniwa and Kazanina2010). These promising results reported in the literature based on adult data raise the question of whether or not children use vowel harmony cues in a similar way in learning words. To investigate this question, Mintz and Walker (Reference Mintz and Walker2006) conducted a head-turn experiment in which infants aged 0;7 listened to nonsense sequences such as detipobubeditopu and observed that infants were sensitive to vowel shifts in such sequences and recognized harmonic sequences such as deti or pobu as words. These findings suggested that when harmonic information was available, children could use it to assign word boundaries. Similar results were reported by van Kampen, Parmaksiz, van de Vijver and Höhle (Reference van Kampen, Parmaksiz, van de Vijver, Höhle, Gavarró and Freitas2008) based on a study where children acquiring Turkish and German were tested using artificially created words. These studies showed that children were capable of segmenting words following vowel cues, but they did not show whether they could use vowel harmony as a segmentation mechanism in word learning in real life.

It is important to note that languages created and tested in the literature are all artificial languages where harmonic cues reliably point to word boundaries. In other words, when word segmentations coincide with non-harmonic vowel shifts, they are attributed to the role of vowel harmony (or disharmony) in segmentation. The question that still remains to be answered is whether natural speech has such cues and thus whether it is reasonable to assume that children acquiring harmonic languages have access to reliable harmonic cues.

In a learning mechanism where vowel harmony is used, words are expected to be consistently harmonic, as a non-harmonic word would give an erroneous cue or a false alarm for a word boundary. Such false alarms are difficult to undo in a mechanism that operates on vowel harmony cues. Similarly, a mechanism that assigns word boundaries only at vowel shifts will have many ‘misses’ due to utterances that are composed of words that have only the same type of vowel. In such instances, a learner would treat the whole utterance as one single word unless s/he can rely on some other cues. Unlike the artificial languages that are used in experiments, natural languages have many false alarm contexts as well as misses, which make it challenging to segment words with vowel harmony cues. Therefore a learning mechanism should assume additional cues, and such cues are available in natural languages in the form of the phonotactic regularities of phonemes, stress patterns, the distributional properties of words, or some other regularities. Earlier studies have already shown that stress patterns were acknowledged as more reliable cues by speakers of Finnish, Turkish, and French, for example, and were preferred over vowel cues (Kabak et al., Reference Kabak, Maniwa and Kazanina2010; Vroomen et al., Reference Vroomen, Tuomainen and Gelder1998). In this study we investigate the nature and the amount of vowel harmony cues that could be used with other cues to segment words in natural speech by contrasting the vowel distributions within and across words.

Because we investigate the relevance of vowel harmony to word segmentation, we need to examine the harmony patterns both within words, and across word boundaries. The motivation for this study lies in the idea that, although harmonic languages are not fully harmonic, and word boundaries are not necessarily non-harmonic, there may still be some contrastive regularity in ‘within word’ vs. ‘across words’ contexts that could potentially complement other segmentation cues. For example, although a harmonic sequence such as /aa/ or /au/ is observed both at word boundaries and within words, it may be more frequent within words. An opposite trend may be recorded for non-harmonic sequences such as /ai/ or /ae/, i.e., they may appear across word boundaries more often. If we can document such a complementary distribution pattern, it may enable us to recognize vowel harmony as a possible segmentation tool. To this end, we look at vowel distributions in two harmonic languages, Turkish and Hungarian. We further contrast the patterns in harmonic languages with the patterns in two non-harmonic languages, Farsi and Polish, through parallel analyses run on the four languages individually. We can assume a mechanism that can rely on vowel harmony only if within- and across-word distributions are reliably different, especially in harmonic languages.

Vowel harmony in Turkish and Hungarian

In both Turkish and Hungarian, words tend to have either all front or all back vowels. Turkish has eight vowels (four front and four back), as seen in Table 1. According to the rules of the front/back or palatal vowel harmony, a word can have either all front or all back vowels (Clements & Sezer, Reference Clements, Sezer, van der Hulst and Smith1982; Demircan, Reference Demircan1996; among others). Turkish word formation is mainly realized through suffixation, and suffixes undergo vowel harmony in such a way that the harmonic status of the words is maintained. In the word at-lar-ımız-dan ‘horse-plu-poss-abl’,‘from our horses’ the word at has a back vowel and all the suffixes that are attached to this word are back (i.e., -lar,-ımız, -dan). In contrast, a word such as ev has a front vowel, and takes the front variety of the suffixes (i.e., -ler, -imiz, -den): ev-ler-imiz-den ‘house-plu-poss-abl’. The language has many non-harmonic words as well due to borrowings and compounding. Some of these non-harmonic words, including the word anne ‘mother’ are very common in child-directed and child speech. Despite these non-harmonic words, a great majority (90%) of words are reported to be harmonic, based on a Turkish corpus study of 601 words (Rodd, Reference Rodd and Ellison1997). In the case of the non-harmonic words, suffixes harmonize in accordance with the last vowel of the word. Some inflections, such as the progressive marker -Iyor, harmonize only partially (it appears as -üyor, -uyor, -iyor, and -ıyor, with the change of the first vowel only; the last vowel remains as /o/). Besides the palatal vowel harmony, Turkish has rounding harmony that applies to high vowels as well, but it will be beyond the scope of this study.

Table 1. Turkish and Hungarian vowels in IPA symbols and orthography

Hungarian has seven short vowels and seven long vowels, which are longer versions of the short counterparts, with two exceptions (see Table 1). The sound represented by é is the long version of /e/, not /ε/ and the sound represented by á is /α:/ not /ɒ:/. In addition to these fourteen vowels, nearly half of Hungarian speakers distinguish an eighth vowel, a short vowel, ë /ε/ as in the word szëg ‘carpenter's nail’ (Abondolo, Reference Abondolo and Comrie1987), but it was not found in the corpus, so it was not covered in the analyses. Just as is the case in Turkish, Hungarian has an internal vowel harmony that concerns the vowel sequences within word stems, and an external vowel harmony through which suffixes harmonize in word formation (e.g., szűr-tök ‘you-plu strain’ vs. szúr-tok ‘you-plu pierce’). An important exception to this generalization is the words that have i and í, so-called ‘neutral’ vowels, which take suffixes with back vowels although they are front (e.g., ír-tok ‘you-plu write’). Just like Turkish, Hungarian has rounding harmony as well (e.g.,-tok/-tök, as exemplified above vs. -tëk after non-round vowels), but it will be beyond the scope of this study.

In general, the terms ‘harmonic’ and ‘non-harmonic’ are used to refer to relations between vowels within words. In the present study it is important to analyze harmony relationships both within words and across word boundaries. For the sake of simplicity, we use these same terms to refer to relations of vowels across word boundaries as well. Those word boundaries that have only back or only front vowels on both sides of the boundary are categorized as harmonic. In an utterance such as Erel-cim bu ne ‘Erel-dim this what’ ‘Erel-dear, what's this?’ from the CHILDES Turkish corpus (File 1aa), the last vowel of the word erelcim is a front vowel. The vowel in the following word is back. The one in the last word is front, resulting in the sequence front–front–front # back # front, where # stands for a word boundary. So a shift from front to back and then back to front vowel corresponds to a word boundary in this particular example. In contrast, in an utterance such as kimin teybi ‘who-gen-3s (tape)recorder-poss-3s’ ‘Whose (tape) recorder?’ from the same corpus, all the vowels are front; therefore there is no word boundary cue provided by vowel shifts.

Farsi and Polish as control languages

In order to test whether the regularities that are observed are due to the harmonic nature of Hungarian and Turkish and are not just accidental, results from harmonic languages are compared to non-harmonic languages. Farsi, an Indo-European language spoken in Iran, is selected as a non-harmonic language because its vowel system is symmetrical – it has three front and three back vowels, as seen in Table 2. In regular orthographic spelling, vowels are not represented, but a corpus with phonologically transcribed utterances (including vowels) is available in CHILDES. In the table, the vowels are shown just as they are represented in the orthography of the corpus.

Table 2. Farsi and Polish vowels in IPA symbols and orthography

Polish is a West Slavonic language of central Europe. It is written in the Latin alphabet with a few additions. Its transcription is transparent in terms of the pronunciation of vowels, which are listed in Table 2 (Jassem, Reference Jassem2003). In addition to three front and three back vowels, it has two nasal vowels, which are marked clearly in orthography. The vowel /u/ has two different representations, u and ó. When i appears before vowels, it is not pronounced as /i/, it rather palatalizes the preceding consonant, so in such sequences it is not treated as a vowel. The alphabet character y represents a variant of /i/ rather than a consonant and is treated as such. It is not included in the vowel chart as a distinct phoneme in some sources (e.g., Stone, Reference Stone and Comrie1987). An important difference between Polish/Farsi and Hungarian/Turkish is that in non-harmonic languages, unlike the harmonic ones, suffixes do not undergo vowel harmony to fit the properties of the vowels in the stems.