【作者】 孙承杰；

【导师】 王晓龙；

【作者基本信息】 哈尔滨工业大学 ， 计算机应用技术， 2008， 博士

【摘要】 随着计算技术和生物技术的进步,当前生物医学文献正在以前所未有的速度增长。这些文献中蕴含着最新的研究进展和丰富的生物医学知识,对于生物医学研究者具有重要意义。然而数以千万计的文献使得研究者追踪和整理自己需要的知识和信息变得越来越困难。文本挖掘技术可以解决这一问题,帮助生物医学研究者提高从文献中获取知识和信息的效率。因此针对生物医学文献的文本挖掘研究具有重要的应用价值。判别式模型是一类直接利用特征来预测目标变量的发生概率的机器学习模型,本文中主要用到的判别式模型有最大熵模型和条件随机域模型。相对于产生式模型,判别式模型降低了特征之间的独立性假设的要求,并且与很多文本挖掘任务的需求相一致,因而更有可能取得好的效果。本文主要研究如何利用判别式模型来解决生物医学文献挖掘中的问题。具体地,我们研究了生物医学文本挖掘中的三个任务:生物医学名实体识别、生物医学实体规范化以及生物医学语义关系抽取。在这3任务中,第二个任务是第一个任务在语义处理上的延伸,前两个任务是第三个任务的基础。本文的主要内容包含以下4个方面。生物医学名实体识别的目标是确定一个给定的文本集合内的某一类型的实体的名字的所有实例,它是进行深层次文本挖掘的必要步骤之一。本文在考察了生物医学领域实体识别的特点和难点,分析了目前已有的生物医学实体识别方法的优缺点的基础上,提出了利用条件随机域模型结合丰富特征集来进行生物医学实体识别的方法。这些特征包括:构词法特征、上下文特征和句法特征。其中,浅层句法特征是首次被引入到条件随机域模型中,同时用来进行实体的边界检测和类别判断。实验表明,这一特征可以有效地提高名实体识别的效果。有监督的机器学习方法需要大规模的标注语料。大量的电子文献使得在生物医学领域获取未标记的语料已相当容易,但是对语料进行标注仍然是一件昂贵的工作。针对在生物医学名实体识别中有监督学习所需的大规模训练语料比较难以获取的问题,本文提出了基于最大熵模型的协同训练的半监督学习方法。该方法可以利用大量的未标注语料来提高在较少的标注语料的基础上学习到的分类器的名实体识别性能。为了进一步提高半监督学习的效果,本文将主动学习引入到半监督学习的过程中。实验表明,基于最大熵模型的协同训练方法可以有效地提高初始分类器的识别性能。灵活的生物医学实体命名方式使得生物医学实体具有严重的歧义。这已成为对生物医学文献进行深层自动文本挖掘的主要障碍之一。生物医学实体规范化的提出就是为了解决这一问题。生物医学实体规范化就是把生物医学文献中表达同一概念的不同变体映射到统一的概念标识符。本文提出了一种用于生物医学实体规范化的多层歧义消解框架。实体规范化过程中不同阶段有不同的歧义情形,在本文提出的框架中,针对这些情形采用了有针对性的解决策略,包括:基于词典的实体名字检测,基于机器学习方法的候选选择以及基于知识的歧义消解。在BioCreAtIvE2006基因名字规范化任务的测试集上的实验表明本文提出的框架可以有效地解决规范化过程中的各种歧义。生物医学语义关系抽取是生物医学文本挖掘的主要研究内容之一,是从无结构的生物医学文献中抽取出生物医学知识的重要手段。在实际应用中,生物医学语义关系的定义有宽泛和具体之分。本文将宽泛定义和具体定义的生物医学语义关系抽取分别看作二分类和多分类问题,提出基于最大熵模型的生物医学语义关系抽取的方法。针对不区分类别的蛋白质相互作用这种宽泛定义的关系抽取,提出了一种基于最大熵的二阶段蛋白质相互作用关系抽取方法。针对多类别的蛋白质相互作用这种具体定义的关系抽取,提出使用最大熵模型结合词特征的抽取方法,该方法在一个具有10种蛋白质相互作用类别的数据集上取得了73.4%的总体精确率。同样的方法应用到疾病与治疗方式关系抽取任务中,也取得了很好的实验结果。此外,本文还通过理论分析和实验对比,从理论和实践两个方面说明了判别式模型比产生式模型更适合生物医学语义关系抽取问题。更多还原

【Abstract】 With the advancement of computing technology and biotechnology, the amountof biomedical literature is increasing in an unprecedented speed. The literature con-tains the latest research progress and rich biomedical knowledge, which are vital forbiomedical researchers. However, tens of millions of literature makes tracking andcollating the necessary knowledge and information become more and more difficult.Text mining technology can solve this problem and enhance the efficiency of utilizingbiomedical literature. So it is valuable in practice to research the text mining tech-nology for biomedical literature. Discriminative models are a class of models usedin machine learning, which can directly use the features to predicate the probabilityof target variables. In this thesis, conditional random fields model and maximum en-tropy model are used. Compare to generative models, discriminative models needn’tthe assumption that features have to be independent and are consistent with the re-quirements of many text mining tasks. So discriminative models are more likely toachieve good results.This thesis is on how to make use of discriminative models to solve the biomed-ical text mining issues. Concretely, we study on three tasks in biomedical text min-ing: biomedical named entity recognition, biomedical named entity normalization andbiomedical semantic relation extraction. In the three tasks, the second is the extensionof the first in semantic processing; the first and the second are the basis for the third.The major contents of this thesis include the following four parts.The target of biomedical named entity recognition is to identify the named en-tity instances of the specified categories in the documents. It is a necessary step fordeep text mining. On the basis of investigating the characteristics and difficulties ofbiomedical named entity recognition and analyzing the advantages and disadvantagesof current methods for biomedical named entity recognition, we propose to use condi-tional random fields model with rich feature sets to identify biomedical named entity.The feature sets include literal feature, context feature and syntactic feature. In thesefeatures, shallow syntactic features are first introduced into conditional random fieldsmodel when doing boundary detection and semantic labeling at the same time, which effectively improve the model’s performance.Supervised machine learning methods need large annotated corpora. Currently,it is easy to obtain un-annotated data in biomedical domain due to the existence ofhuge amount of electronic literature, but corpus annotation is still an expensive work.In order to deal with the lack of large scale annotated biomedical named entity corpus,this thesis proposes maximum entropy based co-training method. This method cantake advantage of the un-annotated data to improve the performance of the classifierstrained on a small scale annotated corpus. Active learning strategy is also integrated tofurther improve the results of co-training. Experiments show the effect of the proposedmethod.The ?exible nomenclature of biomedical named entities results in severe seman-tic ambiguity, which is an obstacle for deep biomedical text mining. Biomedicalnamed entities normalization is an effect way to resolve this problem. The goal ofbiomedical named entities normalization is to correctly associate the named entitiesin documents with standard identifiers. In this thesis, a multi-level disambiguationframework is proposed to accomplish biomedical named entities normalization task.Aiming at different ambiguity situations during the procedure of biomedical namedentities normalization, three different strategies are included in the framework. Theyare dictionary based named entities detection, machine learning based candidate se-lection and knowledge based disambiguation. Experiment results on the test data ofBioCreAtIvE2006 gene name normalization task show that the proposed frameworkcan resolve all kinds of ambiguities during normalization processing effectively.Biomedical semantic relation extraction is one of the main research topics inbiomedical text mining. It is an important mean to extract biomedical knowledge frombiomedical literature. In practice, there are two kinds of relation definition: generaland concrete. The general and concrete definitions are considered as binary classifica-tion and multi-way classification problems respectively and maximum entropy modelis proposed to solve the problems. For a general relation definition, Protein-ProteinInteraction (PPI) relation, we propose a two-phrase PPI Relation extraction methodbased on maximum entropy model. For a concrete relation definition, multi-class PPIrelation, we propose a method which uses maximum entropy model with word fea-tures. In a 10-class PPI relation test data, the method achieved 73.4% accuracy. Thesame method is also applied to a disease-treatment relation extraction and get good re- sults. Besides, we show that discriminative models are more suitable than generativemodels for biomedical semantic relation extraction in both theory and practice.更多还原