The article presents a computational model that predicts functional magnetic resonance imaging (fMRI) neural activation associated with the meanings of concrete nouns, using a combination of a trillion-word text corpus and observed fMRI data for viewing several dozen concrete nouns. The model is trained to predict fMRI activation for thousands of other concrete nouns in the text corpus, achieving highly significant accuracies over 60 nouns for which fMRI data are available. The model's predictions are based on the idea that the neural basis of semantic representation is related to the distributional properties of words in a large text corpus. The authors trained competing computational models based on different assumptions about the underlying features used in the brain for encoding the meanings of concrete objects and found that the best model predicts fMRI activation well enough to successfully match words it has not encountered to their unseen fMRI images. The results establish a direct, predictive relationship between word co-occurrence statistics in text and neural activation associated with thinking about word meanings. The model's predictions are evaluated using cross-validation and show high accuracy, even when predicting words from different semantic categories or within the same category. The learned basis set of fMRI signatures for the 25 semantic features also exhibits commonalities across participants, suggesting that the neural representations of concrete nouns share certain regularities. The research shifts the paradigm for studying neural representations from cataloging fMRI activity to building computational models that predict fMRI activity for arbitrary words.The article presents a computational model that predicts functional magnetic resonance imaging (fMRI) neural activation associated with the meanings of concrete nouns, using a combination of a trillion-word text corpus and observed fMRI data for viewing several dozen concrete nouns. The model is trained to predict fMRI activation for thousands of other concrete nouns in the text corpus, achieving highly significant accuracies over 60 nouns for which fMRI data are available. The model's predictions are based on the idea that the neural basis of semantic representation is related to the distributional properties of words in a large text corpus. The authors trained competing computational models based on different assumptions about the underlying features used in the brain for encoding the meanings of concrete objects and found that the best model predicts fMRI activation well enough to successfully match words it has not encountered to their unseen fMRI images. The results establish a direct, predictive relationship between word co-occurrence statistics in text and neural activation associated with thinking about word meanings. The model's predictions are evaluated using cross-validation and show high accuracy, even when predicting words from different semantic categories or within the same category. The learned basis set of fMRI signatures for the 25 semantic features also exhibits commonalities across participants, suggesting that the neural representations of concrete nouns share certain regularities. The research shifts the paradigm for studying neural representations from cataloging fMRI activity to building computational models that predict fMRI activity for arbitrary words.