.png){kind=icon}
.png){kind=icon}
Return
Semantic-consistent learning for one-shot joint entity and relation extraction
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-03812-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Entity and relation extraction is an essential task in knowledge acquisition and representation. Since novel relations continue to emerge in practice, the extraction for novel relational triplets is urgently needed but understudied. Existing works heavily rely on large-scale artificial external knowledge instead of just the corpus itself to predict novel relations, which require costly human work. In this paper, we propose a semantic-consistent learning method (SCL) for one-shot joint entity and relation extraction, whichbootstraps novel relational triplets from only a single instance of each class. The key is to take full advantage of semantic consistency. Specifically, we estimate the semantic consistency not only between relations and triplets by entropy to refine candidate triplets generated by a joint model, but also between triplets and plain texts by template-based contrastive learning for candidate triplets ranking. Experimental results on benchmarks demonstrate that our method can generalize novel relational triplets with strong robustness and achieve promising performance by micro F1-score 45.12% and 67.02% on two common datasets, respectively.
Keywords:
Joint extraction
One-shot
Semantic consistency
Transfer learning
Journal
IF:
3.5
Papers:
7.5K
Citations:
1.7W