A4 Refereed article in a conference publication
Predicting reaction times in word recognition by unsupervised learning of morphology
Authors: Virpioja S, Lehtonen M, Hulten A, Salmelin R, Lagus K
Editors: Honkela T, Duch W, Girolami M, Kaski S
Publisher: Springer-Verlag
Publication year: 2011
Book title : Artificial Neural Networks and Machine Learning - ICANN 2011
Series title: Lecture Notes in Computer Science
Issue: 6791
ISBN: 978-3-642-21734-0
Web address : https://researchportal.helsinki.fi/en/publications/3d077ce8-4c1e-4a93-bc2a-485ec0d2be17
Abstract
A central question in the study of the mental lexicon is how morphologically complex words are processed. We consider this question from the viewpoint of statistical models of morphology. As an indicator of the mental processing cost in the brain, we use reaction times to words in a visual lexical decision task on Finnish nouns. Statistical correlation between a model and reaction times is employed as a goodness measure of the model. In particular, we study Morfessor, an unsupervised method for learning concatenative morphology. The results for a set of inflected and monomorphemic Finnish nouns reveal that the probabilities given by Morfessor, especially the Categories-MAP version, show considerably higher correlations to the reaction times than simple word statistics such as frequency, morphological family size, or length. These correlations are also higher than when any individual test subject is viewed as a model.
A central question in the study of the mental lexicon is how morphologically complex words are processed. We consider this question from the viewpoint of statistical models of morphology. As an indicator of the mental processing cost in the brain, we use reaction times to words in a visual lexical decision task on Finnish nouns. Statistical correlation between a model and reaction times is employed as a goodness measure of the model. In particular, we study Morfessor, an unsupervised method for learning concatenative morphology. The results for a set of inflected and monomorphemic Finnish nouns reveal that the probabilities given by Morfessor, especially the Categories-MAP version, show considerably higher correlations to the reaction times than simple word statistics such as frequency, morphological family size, or length. These correlations are also higher than when any individual test subject is viewed as a model.
UTU Research Portal