Context Effects in Sentence Reading

The current work investigates the effect of context on the neural representation of concepts, measured by MEG during reading of sentences in active and passive voice. Two approaches to incorporating context are proposed − one by including estimates of the functional connectivity of the MEG sensors, and the other by creating a joint representation of the sentences using an HMM. The suggested methods are evaluated against non-context models in two tasks − classifying sentences as active or passive and identifying pairs of active and passive sentences that exemplify the same concept. Both supervised and unsupervised learning frameworks are presented, as well as the effect of the number of hidden states in the unsupervised framework on the task accuracies.