COMPUTING RESYLLABIFICATION RATES AND INFORMATION CONTENT
2.6. General discussion
This chapter described three related investigations that justify a speech production model with syllabic structural information within the lexicon on computational grounds. The first study on resyllabification quantified the actual resyllabification of connected speech in English, Italian and Hindi. There is often at least some diglossia in many languages and the written variety (especially if taken from standard publications) is not reflective of its everyday use. Actual speech often contains speech disfluencies such as ‘uh’ or ‘um’ that are omitted in written forms and which might potentially resyllabify. Therefore, this analysis of
0
67 speech transcriptions will perhaps encourage others to quantify spoken speech corpora in other languages.
The results show that on average, 33% of syllables in a conversation are resyllabified in English while less than 1% were resyllabified in Italian and Hindi. While the resyllabification rate for English can be considered relatively high, it was also illustrated that this was mostly isolated to a small number of high-frequency syllables. Therefore, the potential savings from resyllabification are quite high.
This picture could change within the context of other languages and their syllable typology. Many agglutinative languages would be intuitively expected to have a higher resyllabification rate. However, we must be careful in making such generalisations as some languages (e.g., Japanese) have a majority of CV syllables and may have a very low resyllabification rate (Itô & Mester, 1995). As Japanese is mora-timed as opposed to stress-timed, syllabification will not be affected by stress assignment. Other languages (e.g., Telugu) almost always have word-final vowels and may therefore only resyllabify occasionally. The study of Dravidian languages might be interesting in this context as the spoken variety tends to encourage word-final vowels while simultaneously having an unusually large number of function words and morphemes that are vowel-initial (Zvelebil, 1990). The implications of this on resyllabification might shed light on how syllable structure is represented cross-linguistically. The main point is that the major speech production models that are in current circulation are based primarily on European and particularly Germanic languages. While some of them provide a good understanding of speech production within their domain, there is a need for a larger sample from other languages. It may be that different languages may store different structural information and the resyllabification rate might be a good indication of whether such information is stored or not.
68 The next study focused on the information content of various syllabic representations within the context of different speech production models. Storing syllable structure may lead to computational benefits, but the trade-off is increased storage costs. The hierarchical structure that needs to be stored on top of the basic phonemic sequence adds to the capacity required by lexical representations, but by how much? Representations that are not syllabified require that the phonemic segments are stored along with their serial order. This is a minimal requirement. Since the phonemic content of words will be common to all accounts, this analysis (initially, at least) takes this to be a constant. The question is how much information is required to store the serial sequence in comparison to what is required by a hierarchical syllabic representation where the order of some positions is predictable based on grammatical principles. The analysis showed that this additional storage is relatively modest even when phonological content is attached to nodes at the bottom of a hierarchy. If phonological content is distributed over the hierarchy (as would be the case for feature-geometric accounts), the storage requirements are reduced still further.
The results showed that there was substantial savings for storing syllable structure as opposed to serialised phonemes (as in the LRM model). This is of course within the context that syllables are represented in the lexicon as abstract grammatical units (i.e., syllables provide the rules of combination for abstract phonemic sequences) and not units that are an embodiment or restatement of more fundamental acoustic or articulatory dimensions.
Finally, the combinatorial advantages in storing syllable structure within the context of a language acquisition system were illustrated. If grammatical principles of combination do not play a major role in language acquisition, a larger potential phase-space needs to be prearranged for storing newly acquired information. If however, the structural information embedded from the start, much less initial phase-space is needed. This final analysis is speculative and assumes that memory is analogous to a content-addressable memory system.
69 We do not use this analysis to justify such a view of human memory, but rather to illustrate how a hierarchical organisation of word-forms (e.g., a syllable hierarchy) could be advantageous in organising the mental lexicon.
This chapter was an attempt to explore the computational consequences for storing abstract syllabic structural information within the mental lexicon. It illustrated the pitfalls of assumed conjectures in areas such as resyllabification and showed how actual quantification is the only means of establishing the syllable’s place within speech production. The results show that when taken together, the resyllabification rate, the information content, and advantages in storage and combinatorial efficiency, storing syllable structure has substantial savings. While this is by no means a replacement for experimental data, it can be taken as a guide for the psycholinguistic experiments in the following chapters that can bring us closer to an accurate (and perhaps universal) speech production model. It is also hoped that this work will encourage psycholinguists (as well as provide them with some initial data) to conduct studies on connected speech as this is a rarely studied, but highly significant, aspect of psycholinguistics.
70