【作者】 王杰；

【导师】 苏新宁；

【作者基本信息】 南京大学 ， 图书情报（专业学位）， 2020， 硕士

【摘要】 近几年,基于被引片段的自动摘要正成为研究的热点,原因在于被引片段具有避免认知价值偏移,与原文表述相一致,更能反映原文信息等优点,并且其在引用动机分析、主题识别等领域中也起着重要的作用。但目前对被引片段的识别研究还不够深入,在CL-Sci Summ共享任务中最高的识别效果为14.5%,该值较低。同时大多数研究采用的是基于特征的机器学习方法,缺点在于需要人工提取重要特征,费时费力,不能有效挖掘引文内容和被引片段之间的关系。因此,为自动提取文本特征,以及进一步提升被引片段的识别效果,本文将其作为文本匹配任务,即根据提供的引文内容从被引文献中获取与之最相似的句子。本文结合深层双向语言模型BERT和因子分解机模型FM,提出深度文本匹配模型BertFM,该模型不仅能将输入的文本信息自动表示为向量形式,还能将文本之间的特征进行二阶组合,以充分获取引文内容和被引片段之间的潜在关联。同时,BertFM在CL-Sci Summ 2019共享任务测试集上的F1值达到25.54%,说明BertFM模型能够有效提高被引片段的识别效果。本文首先对训练数据集中正样本进行整体特征分析,以了解其分布情况。其中,被引片段章节位置分析说明被引片段主要是总结性语句,且一般位于文献和章节的开头或结束部分;序列长度分析表明绝大多数引文内容和被引片段具有16-32个单词,同时长度小于等于128时能够覆盖99%的文本数据;被引片段句子数量分析反映了99.89%的被引片段最多为3句话。由于被引片段识别会出现严重的类别不平衡问题,本文通过相关研究工作,选择随机欠采样和最近邻欠采样方法来平衡正负样本。通过对比实验发现,当使用随机欠采样比例为1时,BertFM能够取得较好的识别效果。在得到使BertFM表现性能较优的训练数据集后,本文在此基础上进行多组对比实验,探究不同因素对BertFM识别效果的影响。(1)文本表示方法。本文在BertFM池化层中采用三种策略来对引文内容和被引片段进行表示:CLS,Mean和Max策略。结果表明,Mean策略能够更加有效的表示文本语义信息,有利于提高识别效果。(2)不同特征。本文探究了章节位置特征,词性特征和词长特征对BertFM模型的影响。结果发现这些新特征均未能改善BertFM性能,原因可能在于BertFM已具备较好的特征提取和表示能力,使得新加入的特征无法有效提供其他信息,甚至会产生冗余信息。(3)不同超参数。BertFM模型中主要的超参数为文本序列长度和FM隐参数。通过多个数值对比实验,本文选取序列长度128,FM隐参数512作为BertFM模型的最终参数。(4)Dropout机制。本文研究了Dropout机制对文本特征交互的影响,结果表明Dropout会丢弃引文内容和被引片段之间一些重要的二阶组合特征,使得最终的识别效果降低。(5)不同句子返回策略。为比较返回句子数量对被引片段识别的影响,本文共设置6种不同的返回策略。实验结果证明,当返回句子数量为2或3时,BertFM模型具有较好的识别效果,且在句子数量为3时具有最高的F1值,其达到25.54%。最后,本文对被引片段识别效果较低的原因进行分析,主要包括引文内容信息不全,被引文献具有缩略词,以及具有语义相似的句子。更多还原

【Abstract】 Currently,automatic summarization based on cited text spans has become an important issue.The reason is that cited text spans has the advantages of avoiding cognitive value deviation,being consistent with original expression and reflecting the original information better.Also,cited text spans plays an important role in citation motivation analysis,topic recognition and other fields.However,the research on cited text spans recognition is not deep enough.The highest recognition result in CLSci Summ sharing task is 14.5%,which is low.Meanwhile,the most scholars adopt feature-based machine learning methods.The main disadvantage is that it requires manual feature extraction,which is time-consuming and labor-intensive,and the relationship between citation content and cited text spans cannot be effectively mined.Therefore,in order to automatically extract text features and improve the performance of cited text spans recognition,we take it as the text matching task,which obtain the most similar sentence from cited references according to citation context.In this paper,we combine deep bidirectional language model BERT and factorization machines model FM to propose a deep text matching model BertFM.This model can not only automatically represent input text information as vector form,but also contain pairwise interactions of text features,which is to get potential links between citation context and cited text spans.Finally,F1 value of BertFM reaches 25.54% on the CL-Sci Summ 2019 shared task test set,indicating that BertFM can effectively improve the result of cited text spans recognition.Firstly,we analyze overall features of positive samples in training dataset to understand data distribution.Chapter position analysis of cited text spans shows that cited text spans are mainly summary sentences and generally located at the beginning or end of paper and chapter.Sequence length analysis indicates that the majority of citation context and cited text spans have 16-32 words,and when the length is less than or equal to 128,it can cover 99% data.Sentence number analysis of cited text spans reflects that 99.89% cited text spans are limited to 3 sentences.Since cited text spans recognition have severely class imbalance problem,we select random undersampling and nearest neighbor undersampling methods to balance positive and negative samples through related work.With the experiments,it is found that when random undersampling ratio is 1,BertFM can achieve better recognition results.After obtaining the training dataset that makes BertFM perform better,multiple group of comparative experiments are conducted to explore the influence of different factors on the recognition impact of BertFM.(1)Text Representation Method.We use 3 strategies in BertFM pooling layer to represent citation context and cited text spans,including CLS,Mean and Max.The results show that Mean strategy can effectively represent text semantic information,which helps improve recognition result.(2)Different Features.This paper explores the effects of chapter position features,part-of-speech features and word-length features on BertFM model.Through the results,we find that none of these new features can improve the performance of BertFM.The reason may be that BertFM already has good feature extraction and representation capability,making added features unable to provide other information and even produce redundant information.(3)Different Hyper-parameters.The main hyper-parameters in BertFM are text sequence length and FM hidden parameters.Based on the experiments,sequence length of 128 and FM hidden parameter of 512 are selected as the final parameters of BertFM.(4)Dropout Mechanism.We study the Dropout impact in feature interactions.The results demonstrate that Dropout can discard some important pairwise feature interactions between citation context and cited text spans,which makes the final recognition effect lower.(5)Different Sentence Return Strategies.In order to compare the impact of sentence return number on cited text spans recognition,a total of 6 different return strategies are set up in this paper.Experimental results prove that when the returned sentences are 2 or 3,BertFM has a better recognition effect.And when the sentence number is 3,it has the highest F1,which reaches 25.54%.In the end,we analyze the reasons of low cited text spans recognition,which include incomplete citation context,cited references with acronyms and semantic similar sentences.更多还原