【作者】 王磊；

【导师】 彭涛；

【作者基本信息】 吉林大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 随着互联网的不断发展与大数据时代的到来,每天都有大量数据的产生,人们需要从这些数据中提取有价值的数据。作为信息抽取的子任务之一,关系抽取通过从数据中提取实体对及其相互关系而广泛应用在搜索引擎等领域中,是大数据时代的重要技术之一。关系触发词作为文本数据中表述实体关系的单词或短语在关系抽取中起重要作用,充分利用其关系特征有助于提高关系抽取的准确率。传统的关系触发词提取方法仅能提取触发某种关系类型的触发词,得到的触发词未必能触发给定实体对的关系。有的关系触发词提取方法能够提取给定实体对的关系触发词,但仅适用于单词型的关系触发词。对于短语型关系触发词,这类方法得到的触发词并不完整。本文利用依存句法分析与词激活力模型,将完整的关系触发词提取划分成核心触发词提取与非核心触发词提取两个阶段,结合两个阶段得到的核心触发词与非核心触发词作为最终的关系触发词。本文提出的方法取得了很好的关系触发词提取结果。在核心触发词提取阶段,使用依存句法分析计算依存距离来评估单词在依存关系图中到实体对的距离,计算序列距离来判断单词在原始单词序列中与实体对的相对位置。结合依存距离与序列距离计算评估分数,根据评估分数与单词词性得到核心触发词。在非核心触发词提取阶段,对词激活力模型进行改进以提高其准确性与解释性,依据单词与核心触发词之间的词激活力与单词词性提取非核心触发词。两个阶段得到触发词一起作为最终的关系触发词。通过搭建Stanford CoreNLP集群提高其文本数据处理效率、建立数据集的词激活力矩阵降低非核心触发词提取的时间复杂度对本文提出的方法进行优化。利用MPI相关技术将本文提出的方法搭建成集群以适应大数据环境,提高大数据下关系触发词提取效率。与传统的触发词提取方法相比,本文提出的方法更充分地考虑触发词与实体对之间的联系以及单词之间的激活关系,得到的关系触发词更准确、更完整。在标注好的SemEval 2010 Task 8数据集以及NYT数据集上进行的实验取得了较好的实验结果,从单词的角度取得了0.87的F1值且在关系实例的角度正确率为72%。实验证明了优化对算法效率的提升及其对大数据环境的适应能力。通过在关系抽取模型中加入关系触发词特征验证了关系触发词对关系抽取的指导作用。将算法得到的关系触发词同给定实体对一起构成三元组保存到Neo4j数据库中并利用echarts技术进行可视化,实验结果直观地体现出关系触发词的准确度及其在开放式关系抽取中的应用。总而言之,本文提出的方法具有一定的学习研究与参考价值。更多还原

【Abstract】 With the continuous development of the Internet and the arrival of the era of big data,a large amount of data is generated every day,and people need to extract valuable data from these data.As one of the sub-tasks of information extraction,relation extraction is widely used in search engines and other fields by extracting entity pairs and their mutual relationships from data,which is one of the important technologies in the era of big data.Relation triggers,as words or phrases that represent entity relationships in text data,play an important role in relation extraction.Making full use of their relation features can help improve the accuracy of relation extraction.The traditional methods of extracting relation trigger words can only extract the trigger words that trigger a certain relation type,and the obtained trigger words may not trigger the relationship of a given entity pair.Some relation trigger word extraction methods can extract relation trigger words for a given entity pair,but only for relation trigger words of the single word type.For phrase-type relation triggers,the relation trigger words obtained by these methods are not complete.This paper uses syntactic dependency analysis and word activation force model to divide the complete relation trigger word extraction into two stages: core trigger word extraction and non-core trigger words extraction.The core trigger word and non-core trigger words obtained from the two stages are combined as the final relation trigger words.The method proposed in this paper has achieved good results of relation trigger words extraction.In the core trigger word extraction stage,the dependency distance is used to evaluate the distance between the word and the entity pair in the dependency graph by using syntactic dependency analysis,and the sequence distance is calculated to determine the relative position of the word and the entity pair in the original word sequence.The evaluation score is calculated by combining the dependency distance and sequence distance,and the core trigger word is obtained based on the evaluation score and the part-of-speech of the word.In the non-core trigger word extraction stage,the word activation force model is improved to improve its accuracy and interpretability,and non-core trigger words are extracted based on the part-of-speech of the word and the word activation force between the word and the core trigger word.The trigger words obtained in two stages are together as the ultimate relation trigger words.The method proposed in this paper is optimized by setting up a Stanford CoreNLP cluster to improve its text data processing efficiency and setting the word activation force matrix of the dataset to reduce the time complexity of non-core trigger words extraction.MPI-related technologies are used to cluster the method proposed in this paper to adapt to the big data environment and improve the efficiency of relation trigger words extraction under big data.Compared with traditional trigger word extraction methods,the proposed method considers the relationships between trigger words and entity pairs and the activation relationships between words more fully,and the resulting relation trigger words are more accurate and complete.The experiments performed on the labeled SemEval 2010 Task 8 dataset and the NYT dataset have obtained good experimental results.From the word perspective,an F1 value of 0.87 is obtained and the accuracy rate is 72% from the perspective of the relation instance.The experiments prove that the optimization improves the performance of the algorithm and its adaptability to the big data environment.By adding the features of relation trigger words to the relation extraction model,the guiding effect of relation trigger words on relation extraction was verified.The relation trigger words obtained by the algorithm are combined with the given entity pairs to form triplets,which are stored in the Neo4 j database and visualized by using echarts technology.The experimental results intuitively reflect the accuracy of the relation trigger words and application of relation trigger words in open relation extraction.In a word,the method proposed in this paper has certain learning and reference value.更多还原