Unsupervised domain adaptation for dialogue summarization with augmented heuristic intermediate domains

Dialogue summarization, an essential component of intelligent dialogue systems, holds significant importance. While end-to-end summarization models exhibit robust performance, their success is contingent upon access to adequately sized annotated datasets. This requirement creates significant acquisition costs, particularly for summarizing dialogues. In this study, we investigate a challenging dialogue summarization scenario where annotated summaries are not available and the availability of data is also limited, broadening the reach and practicality of text summarization techniques. To address this, we propose an unsupervised domain adaptation method for dialogue summarization that leverages rich external, such as news domain data. To address the significant divergences between conversational dialogues–typically characterized by informality, repetition, and multi-party interaction–and standard documents, which are usually formal and succinct, this work proposes a ratio-adaptive mixup approach. This technique enhances data augmentation by generating multiple inferred transitional domains bridging the source and target. With the guidance of heuristic representations and language model discriminator, we develop both source-dominant and target-dominant models exhibit complementary characteristics. Our proposed methods facilitate the gradual transfer of domain knowledge from the source to the target domain. Extensive experiments conducted on three widely-used benchmarks, namely SAMSum, DIALOGSum, and TODSum, demonstrate the superiority of our approach.