在生物医学出版物中评估引用完整性:语料库注释和NLP模型
Maria Janina Sarol1, Shufan Ming2, Shruthan Radhakrishna3
1Informatics Programs, University of Illinois Urbana-Champaign, Champaign, IL 61820, United States.
Bioinformatics (Oxford, England)
|June 26, 2024
概括
研究人员开发了自然语言处理 (NLP) 方法来检测生物医学引文中的微妙引文错误,旨在提高科学完整性. 虽然模型有希望,但准确识别错误的引用仍然具有挑战性.
科学领域:
- 生物医学信息学 生物医学信息学
- 计算语言学 计算语言学
- 学术传播学术交流
背景情况:
- 引用准确性对于科学完整性和评估至关重要.
- 引用中的引文错误可能会扭曲科学证据,人类很难检测到.
- 需要自动化方法来识别生物医学出版物中的引用不准确性.
研究的目的:
- 构建一个生物医学引用的语料库,手动注释的准确性.
- 开发和评估自然语言处理 (NLP) 模型,以识别引用错误.
- 评估不同NLP方法的性能,包括生成的大型语言模型.
主要方法:
- 手动注释3063个引用实例从100个高度引用的生物医学出版物.
- 开发一个NLP管道,结合句子检索 (BM25与MonoT5重排) 和索赔验证 (MultiVerS模型).
- 探索使用GPT-4进行引用准确度分类的少数拍摄的环境学习.
主要成果:
- 大约39.18%的注释引用包含准确性错误.
- 表现最好的NLP模型获得了0.59微F1和0.52宏F1分.
- 与微调模型相比,GPT-4在正确的引用中显示出更高的准确性,但在错误的引用中显示出更低的准确性.
结论:
- 引用引文错误是微妙的,对于当前的NLP模型来说很难发现.
- 开发的NLP模型显示了提高科学文献引用质量和准确性的潜力.
- 创建的语料库和NLP模型是公开的,以促进进一步的研究.
相关概念视频
Improving Translational Accuracy
2.6K
2.6K
Nursing Assessment
7.5K
The two sources for collecting information are primary and secondary. After gathering information, interpretation and validation help to complete the data. The purpose of assessment is to establish data with the initial information, to interpret data about the patient's perceived needs and health problems, and to respond to these problems identified.
The nurse collects all aspects of the patient's health in the initial assessment, establishing priorities for ongoing focused assessments...
The nurse collects all aspects of the patient's health in the initial assessment, establishing priorities for ongoing focused assessments...
7.5K
Pulse Assessment Sites
1.0K
Pulse assessment sites are crucial in evaluating a patient's cardiovascular health. By assessing the pulsations of arteries at specific anatomical locations, healthcare professionals can gather valuable information about blood flow, heart rate, and peripheral circulation. Understanding these pulse assessment sites is essential for conducting comprehensive cardiovascular evaluations and monitoring patients' overall health. These sites are strategically chosen due to the accessibility and...
1.0K
Protein Networks
2.3K
2.3K
Nucleic Acids
7.5K
7.5K
lncRNA - Long Non-coding RNAs
2.8K
2.8K


