Predicting informativeness of words from human brain signals

We study the effect of word informativeness on brain activity associated with reading, i.e. whether the brain processes informative and uninformative words differently. Unlike most studies that investigate the relationship between language and the brain, we do not study linguistic constructs such as syntax or semantics, but informativeness, an attribute statistically computable from text. Here, informativeness is defined as the ability of a word to distinguish the topic to which it is related to. For instance, the word 'Gandhi' is better at distinguishing the topic of India from other topics than the word 'hot'. We utilize Electroencephalography (EEG) data recorded from subjects reading Wikipedia documents of various topics. We report two experiments: 1) a neurophysiological experiment investigating the neural correlates of informativeness and 2) a single-trial Event-Related brain Potential (ERP) classification experiment, in which we predict the word informativeness from brain signals. We show that word informativeness has a significant effect on the P200, P300, and P600 ERP-components. Furthermore, we demonstrate that word informativeness can be predicted from ERPs with a performance better than a random baseline using a Linear Discriminant Analysis (LDA) classifier. Additionally, we present a language model -based statistical model that allows the estimation of word informativeness from a corpus of text