(published in AAAI 2021 ). Key Details from the Research:

: The "AsiaBjBBS" naming convention often appears in large-scale industrial datasets (such as those from companies like Alibaba or Baidu) used to train models to handle real-world service scenarios, such as order status inquiries or technical support.

: Spoken customer service dialogues often contain significant "noise" (redundant phrases, filler words) and common semantics that make standard topic modeling difficult.

: The authors introduce a Topic-augmented Two-stage Dialogue Summarizer (TDS) combined with a Saliency-Aware neural Topic Model (SATM) .

: The model is designed to capture role-specific information (customer vs. agent) and preserve the main ideas of the dialogue in a highly abstractive way.