arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2310.14366cs.CLcs.LG
收藏

基于Bi-Encoders的物种规范化——用于排序的成对句子学习

Bi-Encoders based Species Normalization -- Pairwise Sentence Learning to Rank

Australian Artificial Intelligence Institute(澳大利亚人工智能研究所);UTS(悉尼科技大学);Climate Change Cluster(气候变化集群)

Zainab Awan, Tim Kahlke, Peter Ralph, Paul Kennedy

AI总结:

提出一种基于Bi-Encoders和BERT重排序的深度学习方法,将物种命名实体规范化视为成对学习排序问题,无需特征工程,在LINNAEUS和S800语料库上优于现有技术。

摘要:

动机:生物医学命名实体规范化涉及将生物医学实体与不同的数据库标识符连接起来,以促进跨生物学各个领域的数据整合。现有的生物医学命名实体规范化系统严重依赖字典、人工创建的规则以及高质量的代表特征(如词汇或形态特征)。然而,最近的研究调查了使用基于神经网络的模型来减少对字典、手工规则和特征的依赖。尽管取得了这些进展,但由于缺乏足够大的训练数据集,这些模型的性能仍然受到限制。这些模型倾向于过拟合小型训练语料库,并在面对以前未见的实体时表现出较差的泛化能力,从而需要重新设计规则和特征。贡献:我们提出了一种用于命名实体规范化的新型深度学习方法,将其视为成对学习排序问题。我们的方法利用广泛使用的信息检索算法Best Matching 25来生成候选概念,随后应用来自编码器的双向编码器表示(BERT)对候选列表进行重新排序。值得注意的是,我们的方法消除了特征工程或规则创建的需要。我们在物种实体类型上进行了实验,并使用LINNAEUS和S800生物医学语料库评估了我们的方法与最先进技术的对比。我们提出的方法在将实体链接到NCBI分类方面超越了现有方法。据我们所知,文献中尚无基于神经网络的物种规范化方法。

英文摘要:

Motivation: Biomedical named-entity normalization involves connecting biomedical entities with distinct database identifiers in order to facilitate data integration across various fields of biology. Existing systems for biomedical named entity normalization heavily rely on dictionaries, manually created rules, and high-quality representative features such as lexical or morphological characteristics. However, recent research has investigated the use of neural network-based models to reduce dependence on dictionaries, manually crafted rules, and features. Despite these advancements, the performance of these models is still limited due to the lack of sufficiently large training datasets. These models have a tendency to overfit small training corpora and exhibit poor generalization when faced with previously unseen entities, necessitating the redesign of rules and features. Contribution: We present a novel deep learning approach for named entity normalization, treating it as a pair-wise learning to rank problem. Our method utilizes the widely-used information retrieval algorithm Best Matching 25 to generate candidate concepts, followed by the application of bi-directional encoder representation from the encoder (BERT) to re-rank the candidate list. Notably, our approach eliminates the need for feature-engineering or rule creation. We conduct experiments on species entity types and evaluate our method against state-of-the-art techniques using LINNAEUS and S800 biomedical corpora. Our proposed approach surpasses existing methods in linking entities to the NCBI taxonomy. To the best of our knowledge, there is no existing neural network-based approach for species normalization in the literature.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑