Linguistic Studies of Dialogue

Takao GUNJI, Michinao MATSUI, and Akira OHTANI

Department of Language and Information Sciences,
Graduate School of Language and Culture, Osaka University
1-8 Machikaneyama-cho, Toyonaka, Osaka 560, JAPAN

In this study, we explore the relationship among the syntactic, the morphological, and the phonological components in Japanese.
In doing so, we adopt a variant of constraint-based grammar, which assumes that various kinds of information are shared by many components. The advantage of {constraint-based grammar}, as opposed to {transformational generative grammar}, is that the grammar can express parallel and mutual relationships among linguistic modules and that no directionality in processing need to be assumed. Its grammatical descriptions generally take the form of statements about constraints between features, and about sharing of partial information represented by these features. In short, these are constraints on representations, and are not ones on derivations (i.e., destructive restructuring called transformations).
Since speech sounds are the interface to the outside world, the underlying representation and the surface representation are the complete mirror image of each other both in speech production and speech perception. This property of speech helps us to obtain mutual knowledge of language, which is the advantage of constraint-based grammar. Thus, in this study, we discuss phonological aspects in the framework of constraint-based Japanese grammar.
Two of the important concepts in constraint-based grammar are {feature structures} and {unification} among them. In our approach, we posit two phonological features: the {morph} feature and the {phon} feature. The {morph} feature, which roughly corresponds to the underlying representation, specifies the phonological elements (roughly, phonemes) and their linear ordering. The {phon} feature, on the other hand, roughly corresponds to the surface representation and forms a hierarchical phonological structure. These features can be represented by finite state automata.
This approach, which sets phonological structures in the lexicon, allows them to be independent of syntactic-semantic structures, while at the same time, the correspondence between these two kinds of structures can be represented in the lexicon.
There is one problem in the unification of unconstrained feature structures: when two phonological features don't have a one-to-one correspondence to each other, which often occurs in phonological phenomena, such a structure is in violation of the constraint on unification. Hence, some constraints must exist to avoid such a violation. In our approach, we assume two constraints for this. One is the {Autosegmental Association Constraint} (AAC) and the other is the {Default Value Constraint} (DVC). These constraints play important roles in word formation.
The AAC is originally proposed in the autosegmental theory. This constraint requires roughly that a phonological feature must correspond to another as much as possible. Thus, the AAC produces long vowels, double consonants, euphonic forms ({onbinkei}), tone patterns of words, and so on.
Concerning DVC, the `default' value of the vowel is defined by the sonority hierarchy. It plays an important role in the formation of loan words. The `default' value of the consonant, on the other hand, assumed to be /r/ in Japanese. The feature {C}(onsonant) in the {morph} feature that has no phonological element corresponds to /r/ in the {phon} feature. This hypothesis explains many interesting phenomena about /r/ in Japanese.
Another important aspect in our approach is that we allow variety of candidates for solutions, determining the `optimal' one based on the concept of `cost'. This idea is similar to that in the optimality theory. In this approach, grammatical constraints are deliberately designed to overgenerate possible alternatives for a representation of the given linguistic information. An independent component, which is sensitive to universal and language-specific cost-inducing constraints, will help to choose the most likely candidate. According to this idea, we expect there to be more than one candidates for a phonological feature corresponding to another feature; the candidate that has the minimal degree of violation becomes the ultimate solution.
For example, in producing a sentence, the {phon} feature can freely have any feature values. This means that many candidates can co-exist. But a candidate that keeps better correspondence between the { morph} and the {phon} features is of low cost and can be the ultimate solution. In another case, e.g., word formation, the unification of phonological structures plays an important role in the evaluation of cost.
The study reported here mainly discusses phonological variations (e.g., allophones, double consonants, and so on), word formation systems, and morpho-phonological phenomena (especially verb-inflections in Japanese). We have also developed a formal system of treating accent patterns, which will be further extended in the following years.

Keywords: phonology, morphology, constraint-based grammar, optimality theory, phonological feature structures