PcMSP

我们展示了从305开放获取科学文章中注释多晶材料合成程序 (PcMSP) 的努力，以构建合成动作图。这是一个用于材料科学信息提取的新数据集，同时包含从实验段落中提取的合成句子，以及实体提及和句子内关系。两步人类注释和注释者之间的协议研究保证了PcMSP语料库的高质量。我们介绍了四种自然语言处理任务: 句子分类，命名实体识别，关系分类以及实体和关系的联合提取。综合实验验证了几种最先进的模型对这些挑战的有效性，同时留下了很大的改进空间。我们还进行了错误分析，并指出了一些需要进一步调查的独特挑战。我们将向研究界发布注释方案，语料库和代码，以减轻该领域中标记数据的稀缺性。

We present our effort in annotating Polycrystalline Material Synthesis Procedures (PcMSP) from 305 open-access scientific articles to build synthesis action graphs. This is a novel dataset for materials science information extraction, which contains synthesis sentences extracted from experimental paragraphs, as well as entity mentions and intra-sentence relationships. Two-stage human annotation and inter-annotator agreement studies ensure the high quality of the PcMSP corpus. We introduce four natural language processing tasks: sentence classification, named entity recognition (NER), relation classification, and joint extraction of entities and relations. Comprehensive experiments verify the effectiveness of several state-of-the-art models on these tasks, while leaving significant room for improvement. We also conduct error analysis and point out several unique challenges that require further investigation. We will release the annotation scheme, corpus, and code to the research community to alleviate the scarcity of labeled data in this field.

搜集汇总

背景与挑战

背景概述

PcMSP是一个专注于材料科学信息提取的数据集，包含305篇科学文章中的多晶材料合成程序注释，支持四种NLP任务。该数据集由加州大学圣巴巴拉分校于2022年发布，旨在解决材料科学领域标记数据稀缺的问题。