This dissertation explores two aspects of spoken-word recognition in Japanese: representations of words stored in the lexicon and lexical word competition. The nature of lexical representations and lexical competition were explored by testing three different neighborhood density calculations in naming, word identification in noise and semantic categorization experiments. Neighborhood density is a measure of the number of similar words surrounding a word in the lexicon ("neighbors"). However, definitions of neighbors vary depending on the definition of similarity used. This dissertation tests three neighborhood definitions, each of which coincides with a hypothesis about lexical access with different word representations in Japanese. The first calculation posits the situation where listeners rely on the phonemic word representation as proposed in abstract models. Here, neighborhoods are calculated in terms of the number of phonemes in common, as in the Greenberg-Jenkins calculations (Greenberg-Jenkins, 1964) as widely used in the English word recognition literature.
The second neighborhood calculation included prosodic information as another dimension in the neighborhood calculation in order to reflect the finding that prosodic information has a vital role in Japanese word recognition (Cutler & Otake, 1999). This calculation proposes that Japanese listeners use word-level prosody for lexical access. However, both word representation and word-level prosody are separately stored in the lexicon. In other words, the word representation in this calculation is the categorical abstract representation used in the previous neighborhood calculation and the pitch accent patterns additionally constrain the neighbors. A similarity judgment experiment on pitch accent patterns was carried out and the results were implemented in the calculation.
The third neighborhood calculation is designed to test exemplar-based models. In this calculation, neighborhood density was measured by comparing the similarity of cochleagrams of the 66000 audio files (one file for each noun in the NTT psycholinguistic database, Amano & Kondo, 1999, 2000). Therefore, the word representation is an auditory representation in which all segmental and prosodic information is available. In this calculation, as in the GNM (General Neighborhood Model; Bailey & Hahn, 2000), the words in the lexicon are considered as exemplars and they are mapped onto psychological mental space.
Data for the analyses were collected from Japanese neighborhood experiments using the same 700 test words used in the previous experiments and a lexicon that consisted of only nouns from the NTT psycholinguistic database.
The results of the three experiments in this dissertation shed light on two aspects of lexical access. First, a lexical competition effect is confirmed in Japanese. There are also two types of lexical competition in auditory word recognition: form-based competition (neighborhood density) and phoneme-based competition (cohort reduction). Finally, both abstract (symbolic) representations and episodic (auditory) representations need to be stored in the lexicon. Implications of these results for the current word recognition models are also discussed.