Abstract | ||
---|---|---|
BSTRACTFor summarization in niche domains, data is not enough to fine-tune the large pre-trained model. In order to alleviate the few-shot problem, we design several auxiliary tasks to assist the main task---abstractive summarization. In this paper, we employ BART as the base sequence-to-sequence model and incorporate the main and auxiliary tasks under the multi-task framework. We transform all the tasks in the format of machine reading comprehension [19]. Moreover, we utilize the task-specific adapter to effectively share knowledge across tasks and the adaptive weight mechanism to adjust the contribution of auxiliary tasks to the main task. Experiments show the effectiveness of our method for few-shot datasets. We also propose to firstly pre-train the model on unlabeled datasets, and the methods proposed in this paper can further improve the model performance. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1145/3459637.3482066 | Conference on Information and Knowledge Management |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Qiwei Bi | 1 | 0 | 0.34 |
Haoyuan Li | 2 | 2 | 1.42 |
Hanfang Yang | 3 | 0 | 0.68 |