Due to benefits for teaching and learning, an increasing number of studies have focused on classroom dialogue and how to make it productive. Coding, in which the transcribed conversation is allocated to a set of features, is commonly employed to deal with the textual data arising from this dialogue. This is generally done manually and cannot provide timely feedback to the participants. To address this issue, we explored the possibility of automatically classifying the semantic content of classroom dialogue. Seven categories (prior-known knowledge, analysis, coordination, speculation, uptake, agreement and querying) were distinguished automatically using an artificial neural network-based model. The model achieved acceptable performance and was comparable to human coding. Information about quality of dialogue can be identified in a timely manner. With this knowledge, classroom dialogue can be managed more skilfully, and a more productive form of dialogue is likely to be achieved by teachers and students.