Extraction of Chemical and Drug Named Entities by Ensemble Learning Using Chemical NER Tools Based on Different Extraction Guidelines
Extraction of Chemical and Drug Named Entities by Ensemble Learning Using Chemical NER Tools Based on Different Extraction Guidelines
复制标题
DOI：
--
复制
发表时间：
2015-10
期刊：
影响因子：
--
通讯作者：
T. M. Dieb;Masaharu Yoshioka
中科院分区：
文献类型：
--
作者：
T. M. Dieb;Masaharu Yoshioka
化学命名实体识别是从生物信息学和纳米信息学等多个领域的文本中提取化学信息和化学相关实体(如药物名称和来源材料)的任务。已经有几次尝试根据不同的语料库建设指南来构建处理此类化学相关信息的语料库。尽管这些指南包含常见类型的化学信息，但它们在几个方面有所不同。因此，为特定指南开发的化学NER工具可能能够提取常见的化学命名实体，但它们可能在提取其他与化学相关的实体方面存在问题。假设这些准则之间的差异是一致的，化学NER工具的成功和失败模式也可能是一致的。在本文中，我们提出了一种集成学习方法，它使用条件随机场(CRF)作为一种机器学习技术，根据不同的指南融合各种不同的特征化学NER工具，以构建特定指南的化学NER。为了在这些不同的工具之间实现一致的标记化，我们应用了后标记化机制。我们使用BioCreative IV，CHEMDNER任务数据集对该系统进行了评估。我们证实，组合使用化学NER工具的集成学习方法比仅使用一个化学NER工具的简单领域适应方法更好。我们还证实，集成学习方法可以提高基于规则的化学NER工具在某些任务中的性能。
Chemical named-entity recognition (chemical NER) is the task of extracting chemical information and chemical-related entities such as drug names and source materials from text in several domains such as bioinformatics and nanoinformatics. There have been several attempts to construct corpora for handling such chemical-related information based on different corpus-construction guidelines. Even though these guidelines contain common types of chemical information, they differ in several ways. As a result, chemical NER tools developed for a particular guideline might be able to extract common chemical named entities, but they may have problems extracting other chemical-related entities. Assuming the differences between these guidelines are consistent, the pattern of success and failure of the chemical NER tools might also be consistent. In this paper, we present an ensemble-learning approach that uses the conditional random field (CRF) as a machine-learning technique to fuse a variety of different characteristic chemical NER tools based on different guidelines to construct a chemical NER for a particular guideline. To achieve consistent tokenization across these different tools, we applied a post-tokenization mechanism. We evaluated the system using the BioCreative IV, CHEMDNER task datasets. We confirmed that the ensemble-learning approach using a combination of chemical NER tools is better than a simple domain-adaptation approach using just one chemical NER tool. We also confirmed that the ensemble-learning approach could improve the performance of a well-tuned rule-based chemical NER tool on certain tasks.