Outline 1 Motivation 2 Method 3 Models 4 Results 5 Criticism Murphy, - PowerPoint PPT Presentation

Motivation Method Models Results Criticism Outline 1 Motivation 2 Method 3 Models 4 Results 5 Criticism Murphy, B.; Baroni, M.; Poesio, M. : EEG responds to conceptual stimuli and corpus semantics Course: Mechanisms of meaning Institute for Logic, Language and Computation Paper: EEG responds to conceptual stimuli and corpus semantics

Motivation Method Models Results Criticism Motivation Brain imaging in linguistics Might help isolating semantic aspects from others fMRI (functional magnetic resonance imaging) measures blood flow in brain regions high spatial resolution (millimeters) low temporal resolution (seconds) high costs EEG (electroencephalography) measures electric currents on the scalp low spacial resolution (centimeters) high temporal resolution (milliseconds) moderate costs Course: Mechanisms of meaning Institute for Logic, Language and Computation Paper: EEG responds to conceptual stimuli and corpus semantics

Motivation Method Models Results Criticism Method Subjects are confronted with pictures of animals tools Task is to name them Brain activity is recorded Corpus models are trained with most of the data Models are used to predict for remaining data whether an animal or a tool was viewed (between categories) which animal or tool was viewed (within categories) Course: Mechanisms of meaning Institute for Logic, Language and Computation Paper: EEG responds to conceptual stimuli and corpus semantics

Motivation Method Models Results Criticism Models Search engine (Yahoo!) only allows for searching co-occurrence huge Newspaper (la repubblica) different feature sets window co-occurrence in sentences position discriminates position of verb wrt noun dependency filter filters by paths dependency path paths as features comparatively small Course: Mechanisms of meaning Institute for Logic, Language and Computation Paper: EEG responds to conceptual stimuli and corpus semantics

Motivation Method Models Results Criticism Results The Yahoo! and the window model render results over chance The worse performance of the more sophisticated corpus models is attributed to their sparseness Between categories predictions were more accurate than within categories involved concepts are more similar all concepts were learned with the same feature sets Course: Mechanisms of meaning Institute for Logic, Language and Computation Paper: EEG responds to conceptual stimuli and corpus semantics

Motivation Method Models Results Criticism Criticism Not convinced how this rules out non-linguistic associations Shouldn’t those patterns be cross subject? Training 58 data sets and predicting 2 seems to be a clear case of overfitting Course: Mechanisms of meaning Institute for Logic, Language and Computation Paper: EEG responds to conceptual stimuli and corpus semantics