Combining feature norms and text data with topic models

Many psychological theories of semantic cognition assume that concepts are represented by features. The empirical procedures used to elicit features from humans rely on explicit human judgments which limit the scope of such representations. An alternative computational framework for semantic cognition that does not rely on explicit human judgment is based on the statistical analysis of large text collections. In the topic modeling approach, documents are represented as a mixture of learned topics where each topic is represented as a probability distribution over words. We propose feature-topic models, where each document is represented by a mixture of learned topics as well as predefined topics that are derived from feature norms. Results indicate that this model leads to systematic improvements in generalization tasks. We show that the learned topics in the model play in an important role in the generalization performance by including words that are not part of current feature norms.