Dialogue Text Summarization Method Combined Self-supervised Learning and Neural Architecture Search

Dialogue summary generation task aims at generating summary for dialogue. Due to the changing topic and discrete semantic of dialogue, previous summarization methods on article cannot generate satisfactory summary for dialogue. To reduce the impact of discrete topics on dialogue summary generation, this work proposes a topic division method to form topic paragraph sets. In order to generate dialogue summaries according to the divided topic paragraphs, we further propose a dialogue text summarization model based on a Generative Adversarial framework. In this model, we use the differentiable Neural Architecture Search method to implement the entire search process of generative and discriminative network. The generative network is a BART network formed by sharing the Transformer structure obtained by the search, which we call EBART. The effectiveness of this work is verified by experiments on public dialogue datasets.