Knowledge regarding social information is commonly believed to be derived from sources such as formal relationships and interviews and can be plotted as complex networks. We explored whether social networks can also be extracted through other means by using language statistics. In three computational studies we computed first-order and higher-order (latent semantic analysis) co-occurrences of story characters in three novels. These statistical linguistic frequencies entered in a multidimensional scaling analysis yielded a two-dimensional solution that correlated with the two-dimensional networks of characters generated by experts. An experimental study in which participants were asked to estimate social networks showed that human estimates are similar to computational estimates. These results demonstrated that language statistics based on texts can be used to generate social networks.