Generating One-Turn Dialogue from Given Keywords Based on GPT-2(Chinese) for Oral Chinese Teaching

Pedagogical materials are the basis for carrying out teaching and learning activities. Compared to inefficient manual organization or retrieving methods limited in data scale, text generation technique, aiming to generate readable text, has shown promising potential in automatic generation of pedagogical materials. For purpose of satisfying the need that vocabulary teaching in oral Chinese teaching should expose to more relevant dialogue pedagogical materials, this work builds a model focusing on generating one-turn dialogue from given keywords. The model is built under fine-tuning paradigm and large-scale pre-trained generative language model, GPT-2(Chinese), both of which have not been yet explored in related work, topic-to-essay (TEG) and poetry generation. And this work introduces dialogue role embeddings technique from dialogue system to model topic transfer in one-turn dialogues. Evaluation results show that our model can achieve sufficient quality and diversity simultaneously with comparative novelty in general text generation ability and have the ability to generate one-turn dialogue from given keywords for oral Chinese teaching. Case study hardens the evaluation results. The results provide implications in generating multi-turn dialogue from given keywords.