Authors: Yusen Zhang Ansong Ni Tao Yu Rui Zhang Chenguang Zhu Budhaditya Deb Asli Celikyilmaz Ahmed H. Awadallah Dragomir Radev
Paper reference: https://arxiv.org/pdf/2109.04609.pdf

Contribution

This paper is about an exploratory study on long dialogue summarization and provides several strategies for future works.

Details

Challenge in Long Dialogue Summarization

(1) Effectively use the current neural summarization models on dialogues that greatly exceed their length limits.
(2) Dialogues are interactive, which makes it more context-dependent and the information in dialogues is more sparsely distributed over the text.
(3) Language in dialogue is more informal, which leads to difficulties in modeling relevance and salience.

Current Strategies for lengthy inputs

• Retrieve-then-summarize Pipeline. Retrievers includes TF-IDF, BM25, Locator. etc. Suggestion: develop better utterance retrieval method. Experiments show that the two-step pipeline is better than end-to-end models.
• End-to-end Summarization Models. BART (truncate inputs), HMNet (a hierarchical network for dialogue summarization, with a token level and a turn level encoders), Longformer (16k maximum). Hierarchical network performs better for longer inputs. Pretraining on external dataset may help the performance, but it should be chosen carefully.