Improving Faithfulness in Abstractive Summarization with Contrast Candidate Generation and Selection
摘要中所有实体在source中出现,作为positive sample
we sample examples from the XSum training set where all entities in the ground truth summary appear in the source document.
用相同语义类型,但是不同实体的进行替换,产生不忠实的变体,作为负样本
produce unfaithful variants from the ground truth summary by replacing entities with others that have the same semantic type but different surface form in the source text
pair(pos,neg)进行正负样本的计算
Encode, Tag, Realize: High-Precision Text Editing
1.给出一种新型生成思路,应用任务包括:sentence-fusion,sentence-spliting,abstract-summary
通过标签的delete/keep/P/swap/prounoun等标记后续对token的处理方式
2.tag输出包括 ffd方式和ar两种方式;
3.优势:target set少,依赖于更少的语料,
4.训练数据准备,通过生成前和生成后,能够映射的即为tag-keep,在P集合内的即为keep,其余的记为delete;
P集合准备,通过训练数据准备样式,进行预挖掘准备
5.应用:预处理模块由规则升级为模型
<