New insights into the speaker’s lexicon: Computational modelling and experimental data

点击翻译按钮获取中文摘要

The frequency with which we experience events shapes our subsequent processing of similar events. Language is no exception. In the study of language production, one of the first experimental effects to be discovered was the lexical frequency effect. More frequent words can be accessed and encoded more quickly. This empirical effect has shaped our understanding of the mental lexicon - how information about words is organized in memory- and of the processes involved in accessing these representations. It has been taken as evidence that words are represented in long-term memory together with information about their frequency of use. Just like some words are more frequent than others, some word sequences or utterances are more frequent than others. We say I’ll take a glass more often than I’ll take a plate and a great trip more often than a great plane. Does the frequency of the chunk impact the preparation (i.e., encoding) of this chunk? Current psycholinguistic models of language production predict that this should not be the case. In these models phrases are built on the fly. Once a given utterance has been produced, it does not leave any trace in the system. Yet recent studies have shown that similar effects arise for multi-word utterances. The phrases that speakers use more often can be encoded more quickly. This empirical result demonstrates that the language production system keeps track not only of its encounters with individual words, but also of how often these words are combined into phrases. This result requires that psycholinguistic models of language production be refined or that some of their fundamental assumptions be revisited. The present project capitalizes on this discovery. Its aim is to implement and test a novel model of language production that can account for these effects. This model views the lexicon as a network of interconnected lexical representations that evolves dynamically as a results of the speaker’s use of the network. Tools of Artificial Intelligence (i.e., linear modeling, naïve discrimination learning) will be used to simulate different versions of that model as well as an alternative model in which the system stores representations for whole phrases. Psycholinguistic experiments will be conduct to measure changes in the participants’ performance as a function of their use of the language. The outcome of these experiments will be compared to the predictions of the computational models. This project is expected to mark a turning point in the study of language production processes. It will shed a new light on the cognitive architecture of the language production system and open up new perspectives in the study of that system.