【作者】 王亮；

【导师】 朱巧明；

【作者基本信息】 苏州大学 ， 软件工程， 2023， 硕士

【摘要】 作为自然语言处理(Natural Language Processing,NLP)领域里关系抽取的重要组成部分,事件时序关系抽取在医学、金融和军事等领域有着广泛应用前景,也对诸多需要深入理解文本时序信息的下游自然语言处理任务起着关键作用,例如问答、时间线构建和文本摘要等。事件时序关系抽取的已有研究通常受数据稀疏和不平衡问题的制约,以及其所衍生的模型过拟合问题带来的性能下降。基于上述问题,本文为了从有限的语料资源中挖掘时序关系信息开展研究,具体如下:(1)基于负样本语料再标注的事件时序关系抽取方法针对语料库样本的分布不平衡和语料标注的歧义问题,本文提出一种基于负样本语料再标注的时序关系抽取方法。区别于现有的单标签(Single-Label)时序关系,该方法将时序关系类别重新定义为多标签(Multi-Label)的形式,使得歧义的时序关系能够相互兼容。此外,选取负类样本进行细粒度再标注,可从原本歧义现象严重且占比较大的负样本中挖掘丰富的时序信息。鉴于本方法提供的样本标签与过去研究差异较大,本文进一步提出多标签-单标签映射方法与额外标签映射方法,以充分利用标注的信息。实验结果表明,该标注体系和映射方法可显著提升事件时序关系抽取的性能。(2)基于DCT中心的时序关系抽取方法针对已有的时序关系抽取研究大多仅考虑事件-事件(Event-Event,E-E)间时序关系,忽略事件-时间表达式(Event-Timex,E-T)和事件-文档创建时间(Event-DCT,E-D)间时序关系的问题,本文提出一种基于DCT中心的时序关系抽取方法。该方法围绕DCT这一核心的时序线索,利用多任务学习框架,将多类时序关系抽取任务有机结合起来,最大限度地利用不同类别时序关系标注数据。模型将各类时序关系抽取任务置于统一的神经网络框架之中,利用不同任务带来的更多训练数据以及DCT连接整个文档的桥梁作用,贯穿整个训练和推理过程,为事件-事件时序关系抽取这一主任务提供更为直接、高效的时序线索。在多个任务上进行的一系列实验结果证明了该方法性能层面的优越性。(3)基于时间锚定和负样本降噪的事件时序关系抽取方法针对已有的事件时序关系抽取方法与相关NLP任务的依赖和高耦合问题,本文提出一种基于时间锚定和负样本降噪的事件时序关系抽取方法。具体而言,本文基于时序关系定义设计了一个锚定损失函数,在多任务框架下借鉴文档中时间表达式和DCT固有的时间值来训练事件的“开始-结束”时间。此外,为减少负类样本对时间锚定系统的干扰,提出了一种负样本降噪机制,其能够在整个训练过程中动态地调度模型对负样本学习的权重。实验结果表明,该方法能够有效缓解数据不平衡问题,提升多个时序关系抽取任务的性能。本文针对事件时序关系抽取展开了研究,针对现存的多个问题提出了有效的模型与方法,提升了该任务的性能,可为将来的研究提供参考。更多还原

【Abstract】 As a crucial component of relation extraction in the field of Natural Language Processing(NLP),event temporal relation extraction has broad application prospects in many fields,such as medicine and finance.It also benefits lots of downstream NLP tasks that require deep understanding on the temporal information of natural texts,such as question answering,timeline construction and summarization.Previous studies often suffer from the issues of data scarcity and imbalanced instances,which may cause over-fitting in model training step and bring performance decrease.To relieve above issues and mine more temporal information from limited resources,this dissertation has carried out research work as follows:(1)Event temporal relation extraction based on negative sample re-annotationThe issues of the imbalanced distribution of instances,and the ambiguity of vague instances in existing corpora have severe performance impact on event temporal relation extraction system.To relieve the issues,this dissertation proposes a negative sample reannotation method.Different from previous studies that use single-label temporal relation,this method define the relation as multi-label form,making ambiguous relations be compatible.Besides,our fine-grained re-annotation on the negative samples helps exploit more definite temporal clues from the vague instances.Due to the huge difference between our proposal and previous annotation,we propose two corpus usage strategies,i.e.,multi-to-single and extra label mapping methods,aiming to fully use the annotated information.Experimental results show that our models on the annotation scheme and mapping methods outperform previous work significantly.(2)DCT-centered temporal relation extractionPrevious studies on temporal relation extraction mostly only focus on event-event relation,while the temporal relations of event-timex and event-DCT are often ignored.To resolve this problem,this dissertation proposes a DCT-centered temporal relation extraction method.The method centers on the core temporal clue of DCT,and utilizes multi-task learning framework,which combines several temporal relation extraction tasks into a unified DCT-centered neural network to make the best use of the limited training resource.This method takes advantage of different type of samples,moreover,DCT can act as a hub to connect the whole document,providing more direct temporal clues for the main event-event temporal relation extraction task throughout the training and inference processes.Experimental results show that the proposed method outperforms previous work significantly.(3)Event temporal relation extraction based on time anchoring and negative denoising.Previous neural network models often depend on related NLP tasks,causing high coupling and complexity.To address this issues,this dissertation proposes an efficient event temporal relation extraction model based on time anchoring and negative denoising.In detail,according to the definition of the temporal relations,we design a novel anchor loss for time anchor,besides,we make events learn from the value of time expressions and DCT of the document in a multi-task learning framework.Moreover,in order to decrease the interference of vague instances in training step,we propose a negative denoising mechanism that dynamically adapts the learning weight of negative samples during training process.Experimental results show that our proposal can alleviate the problem of data imbalance and provide performance gains for multiple temporal relation extraction tasks.This dissertation focuses on the event temporal relation extraction,and proposes several efficient methods to improve the task performance and provide reference for future research.更多还原