Lukas, Randi, and Peta will attend and present two posters at Vocab@Leuven this July:
Multimodal modelling of neurological vocabulary representation
How do we internally represent learned vocabulary? We know that processing words in different modalities involves different brain regions – or at least arouses more activity in some than others. But how do their representations interact with each other? Using fMRI data, we investigate how modalities coactivate each other. We particularly focus on three aspects of vocabulary: Orthography, phonology, and semantics.
We will present a computational model based on similarities. Words are sorted into a high-dimensional space based on orthographic, phonological and semantic (dis-)similarity. The model is evaluated on its ability to predict neural response patterns in a passive vocabulary task.
We hypothesize that the activation pattern of one activity can be predicted through the pattern of a different modality of the same word. Furthermore, the representation of a new word could be approximated from the representations of similar known words. We also investigate whether the similarity and consistency of internal representations across modalities relates to strength of recall in a post-test.
Our results will have implications for multi-modal vocabulary learning and provide insight into how newly learned words map to existing representations. We also discuss how our computational model can add to intelligent tutoring systems.
Tracking gradient descent changes during initial second language word learning
The so-called testing effect describes better learning outcomes when trying to retrieve the correct answer as compared to merely restudying the material. The search set hypothesis suggests that during retrieval, we select the answer from a memory search set of possible answer candidates. The aim of this study is to understand the development of this search set during the initial stages of new second language (L2) word learning.
We will investigate this with a retrieval practice learning task. Participants will learn 30 Dutch-Swahili word pairs with 8 learning trials for each word pair. During the retrieval phase, the participants will (make attempts to) type the correct translation of a written word and rate how good they think their answer is. Afterwards they will receive feedback in the form of the correct written translation. Half of the word pairs will be translated from L1 to L2, the other half from L2 to L1. Following the search set hypothesis, we assume that the set of competitors becomes more and more specific, with the competitors gradually becoming more similar to the correct answer. We hypothesize that this will be reflected in the error pattern: When retrieving the new L2 word, the orthographic edit distance between the answers and the correct L1 word should gradually become smaller. When retrieving the existing L1 word, the semantic overlap between the answers and the correct L2 word should gradually become bigger. This effect should be modulated by neighborhood density and also show in the participant’s subjective rating of their answer.
Indirectly, this study will inform us about how a new L2 word form representation maps onto an already existing L1 word form and semantic representation. This knowledge will be used to design model based educational software for improving L2 word learning.