The ubiquity of Internet chat applications has benefited many different segments of society. It also creates opportunities for criminal enterprise, terrorism, and espionage. This thesis proposes statistical Natural Language Processing (NLP) methods for creating systems that would detect the topic of chat in support of larger NLP goals such as information retrieval, text classification and illicit activity detection. We propose a novel method for determining the topic of chat discourse. We trained Latent Dirichlet Allocation (LDA) models on source documents and then used inferred topic distributions as feature vectors for a Support Vector Machine (SVM) classification system. We constructed LDA models in three ways: We considered the collective posts of authors as documents, hypothesizing that we could detect the topic physics given only one side of the conversation. The resultant classifiers obtained F-scores of 0.906. Next, we considered individual posts as documents, hypothesizing we could detect physics posts. The resultant classifiers obtained F-scores of 0.481. Finally, we considered physics textbook paragraphs as documents, hypothesizing that we could determine the topic of an author or a post based on an LDA model created from a textbook and a sample of noisy chat. The resultant classifiers obtained F-scores of 0.848 and 0.536 respectively.