SciERC

我们在科学文章中引入了识别实体、关系和共指集群的多任务设置。我们创建了SCIERC，这是一个包含所有三个任务的注释的数据集，并开发了一个名为SCIIE的统一框架，具有共享跨度表示。多任务设置减少了任务之间的级联错误，并通过共指链接利用了跨句关系。实验表明，我们的多任务模型在科学信息提取方面优于以前的模型，而无需使用任何特定领域的特征。我们进一步表明，该框架支持构建科学知识图谱，我们用它来分析信息非科学文献。

We introduce a multi-task setting for identifying entities, relations and coreference clusters in scientific articles. We create SCIERC, a dataset annotated for all three tasks, and develop a unified framework named SCIIE that adopts shared span representations. The multi-task setting reduces cascading errors across tasks and leverages cross-sentence relations via coreference links. Experiments demonstrate that our multi-task model outperforms prior models in scientific information extraction without employing any domain-specific features. We further show that the framework supports the construction of scientific knowledge graphs, which we use to analyze information in non-scientific literature.