Alan H. Kawamoto

A Neural Network Model of Multistable Perception

Abstract The major properties of multistable perception are reviewed and previous models to explain it are described. We then suggest a neural network model based on state vectors, feedback and Hebbian synaptic modification which can be made to account for a number of the experimental findings. Some simple extensions are made to previous versions of the neural model to allow the dynamic modification of synaptic connectivities during the course of the stimulus presentation. This enables such properties as the time course of reversals, adaptation, and hysteresis to be simulated. In addition, we present evidence demonstrating the importance of both the stimulus and the recent and distant history of the system to disambiguate ambiguous stimuli.

When Two Meanings Are Better Than One: Modeling the Ambiguity Advantage Using a Recurrent Distributed Network

Ambiguous words are processed more quickly than unambiguous words in a lexical decision task despite the fact that each sense of an ambiguous word is less frequent than the single sense of unambiguous words of equal frequency or familiarity. In this computer simulation study, we examined the effects of different assumptions of a fully recurrent connectionist model in accounting for this processing advantage for ambiguous words. We argue that the ambiguity advantage effect can be accounted for by distributed models if (a) the least mean square (LMS) error-correction algorithm rather than the Hebbian algorithm is used in training the network and (b) activation of the units representing the spelling rather than the meaning is used to index word recognition times.

Pronunciation of homographs

Abstract We tested the predictions concerning the pronunciation of monosyllabic homographs (words such as bass that have one regular and one irregular pronunciation) made by the dual-route model, the analogy model, and the distributed model if some level of independence between processes is assumed. In Experiment 1, we found that the naming latency of monosyllabic homographs was longer than the naming latency of regular control words that were half the printed frequency of the homographs. (We also found a longer naming latency for multisyllabic homographs that differed in both stress assignment and phonemic composition (e.g., project ) compared to control words, but no difference in naming latency for multisyllabic homographs that differed only in stress assignment (e.g., insult )). In addition, we found that the more frequent pronunciation was the irregular pronunciation for the majority of the monosyllabic homographs, and the naming latency for the more frequent pronunciation for these homographs was the same as or longer than the naming latency for the less frequent pronunciation. In Experiment 2, we carried out a delayed naming control experiment and did not find any naming latency differences between homographs and controls. In Experiment 3, we compared homographs with exceptions matched in frequency and found that the naming latency difference between homographs and their controls was larger than that between exceptions and their controls. Thus, the longer naming latency for monosyllabic homographs found in Experiment 1 is not simply a production effect or an exception effect. Finally, in Experiment 4, we found that for all three classes of homographs, the proportion of a given pronunciation was highly correlated with the subjective familiarity of that pronunciation. In the discussion, we argue that these results can only be supported by naming models in which the entire input string dominates sublexical constraints. Moreover, we argue that the counterintuitive data in which the more frequent pronunciation has a longer latency than the less frequent pronunciation requires two different constraints or processes that have different time courses.

Distributed Representations of Ambiguous Words and Their Resolution in a Connectionist Network

Publisher Summary This chapter shows how a distributed representation of ambiguous words captures the effects of frequency and contextual biases quite naturally. Moreover, the flexibility of this type of representation allows both polysemy and ambiguity to be treated identically, and suggests how neologisms can be handled. Although the use of distributed representations in psychology is not very widespread, these ideas are continually explored because of the demonstrations of their utility in a variety of different domains. The chapter also explores the effects following the resolution of the ambiguity. When asked, people are able to generate more than one meaning of an ambiguous word. The alternative meaning of an ambiguous word can be accessed by habituating the connections. Consistent with empirical findings, access of the alternative meaning is more likely when the subordinate meaning is given initially.

Runword: An IBM-PC software package for the collection and acoustic analysis of speeded naming responses

Researchers generally collect two dependent measures from most types of speeded responses: reaction time and error proportion. Recently, experimenters have tested lexical theories using alternative empirical measures, such as response force in decision tasks and response duration in naming tasks. We offer a set of software tools that expands the battery of dependent measures normally available to naming experimenters to include various duration and intensity measures of digitized voice recordings. This article explains the functionality, programming logic, and theoretical motivation behind each measure, as well as the software to support digitized naming experiments. The software runs in DOS on IBM-PC-compatible hardware with SoundBlaster 16-bit sound cards, and it takes advantage of the decreasing costs of hard-drive space and digital sound cards.

Time course of regular and irregular pronunciations

Abstract In the past few years, there has been a great deal of discussion over the necessity of two processes instead of a single process to account for both regular and irregular forms. We describe some counter-intuitive naming data (the less frequent pronunciation has a shorter mean latency than the more frequent pronunciation) that cannot be accounted for by Seidenberg & McClelland who use a single measure (the phonological error score) to index both response likelihood and response time. We present two possible scenarios in which feed-forward networks could account for the data, one based on the variance of the distribution of the phonological error scores and the other based on the time course of activation. In this study, we examine the second possibility in a series of simulations using three different feed-forward networks, but find no evidence that feed-forward networks can account for the counter-intuitive data.

The return of “visiting relatives”: Pragmatic effects in sentence processing

Two experiments were performed to investigate the role of syntactic and pragmatic cues on the disambiguation of noun phrases of the form VERB+ing NOUN+s, like visiting relatives, that can be interpreted as either singular or plural noun phrases. Both experiments used a self-paced reading task in which reading times were measured for two words, a verb and an adverb, immediately following the potentially ambiguous noun phrase. The interpretation of the noun phrase as singular or plural was biased by pragmatic cues in the first experiment and by syntactic cues in the second experiment. In both experiments, subjects were faster to read the adverb following the verb when the interpretation biased by the cues agreed in number with the verb that immediately followed the target noun phrase than when it did not agree with the verb. These results suggest that pragmatic cues, like syntactic cues, can be utilized rapidly in sentence processing.

The segment as the minimal planning unit in speech production: Evidence based on absolute response latencies

A minimal amount of information about a word must be phonologically and phonetically encoded before a person can begin to utter that word. Most researchers assume that the minimum is the complete word or possibly the initial syllable. However, there is some evidence that the initial segment is sufficient based on longer durations when the initial segment is primed. In two experiments in which the initial segment of a monosyllabic word is primed or not primed, we present additional evidence based on very short absolute response times determined on the basis of acoustic and articulatory onset relative to presentation of the complete target. We argue that the previous failures to find very short absolute response times when the initial segment is primed are due in part to the exclusive use of acoustic onset as a measure of response latency, the exclusion of responses with very short acoustic latencies, the manner of articulation of the initial segment (i.e., plosive vs. nonplosive), and individual differences. Theoretical implications of the segment as the minimal planning unit are considered.

Locus of semantic priming effects in speeded naming

The locus of semantic priming effects was examined by measuring onset and rime durations as well as response latencies of words with consistent and inconsistent pronunciations, using the postvocalic naming task. We found that the effect of a semantic prime on naming duration was localized rather than spread across the entire word; onset durations were shorter in the related condition than in the unrelated condition, but rime durations were equal in the two prime conditions. Moreover, the priming effect on onset durations was larger for words with inconsistent than for those with consistent pronunciations. These duration results cannot be accounted for by previous proposals, but they can be accounted for by models in which phonemes are activated in parallel rather than serially from left to right and in which motor programs are based on phonemes rather than syllables. Contrary to previous reports of an interaction of prime and regularity (a factor closely related to consistency) on naming latency, we found no interaction of prime and consistency on response latency. We argue that this conflict is only apparent and arises because naming latency conflates response latency and initial phoneme duration for some targets.

Non-Obligatory Vowel Epenthesis in -Ed Pseudowords: Ambiguity in Syllabification Resolved by Syntax and Suffixation

In this study, we consider a number of factors that affect whether psuedowords ending with the suffix -ed are pronounced with an additional syllable. The additional syllable results if vowel epenthesis occurs. Because this process is optional unless -ed is preceded by a dental, an ambiguity in syllabification arises in many cases. in Experiment 1, we found that subjects do occasionally pronounce -ed pseudowords with an additional syllable, although vowel epenthesis for this class of words has been considered to be unproductive (Marchand, 1969). Moreover, as with actual English words, this process occurred when the pseudowords were used in sentences as adjectives, but did not occur when the same pseudo-words were used as verbs. in Experiment 2, we found that addition of the suffixes -ly and -ness increased the likelihood of vowel epenthesis despite the fact that these suffixes generally do not change the pronunciation of the root in Experiment 3, we showed that perceived grammatical category also affected vowel epenthesis; vowel epenthesis was more likely when the pseudowords were used in a preceding sentence with suffixes that adjectives normally take (-ly or -ness) compared to a condition in which the pseudowords were used in an inflectional form of a verb (-es or -ing instead of -ed). The final experiment was a control experiment showing that the results in Experiment 3 were not simply due to priming effects. The results of the experiments on the pronunciation of pseudowords reported here and elsewhere by other investigators suggest that grammatical category information must be included in any complete model of reading aloud. A parallel distributed processing approach that incorporates a grammatical category constraint is discussed.