Building Topic Models to Predict Author Attributes from Twitter Messages

We use the topic modeling software package MALLET [10] to construct models of 100 topics each for the four languages in the scope of the PAN’15 Author Profiling task. The topics in these models are essentially groups of words that may be semantically related and are frequently observed near each other in a collection of training documents. To ensure we had a sufficiently large body of examples to build such models, we collected our own corpora of Twitter messages in English, Spanish, Italian and Dutch. We also use MALLET to infer the most likely distribution over the generated topics that could have produced any given tweet instance, allowing us to represent tweets as concise 100-element document-topic distribution vectors. These representations serve as inputs to a set of classifiers that make predictions for unknown authors’ age, gender, extroversion, stability, agreeableness, conscientiousness, and openness.