随着学者们对科学发展规律探索的不断深入，研究科学系统的科学也逐渐受到更多学界关注。其中科学引文反映了科研成果的传播过程，也可以辅助衡量科研成果的影响力。目前多数对于引用的研究默认这些引用是积极正面的，但实际上也存在着一些带有负面倾向的引用关系，这些负面引用对于科学论文影响力的衡量也有着不可忽视的作用。探索论文之间通过负面引用关系进行相互作用而涌现出的性质特点，对于科学系统的研究也有着重要意义。
本文基于AAN引文语料和ACL ARC引用关系数据分别进行分析，采用情感识别算法计算情感得分，对比分析科学引文与日常评论的负面表达的差别，分析负面引用的文本性质；运用引用行为发生的时间数据，计算相关统计指标随时间的变化，探索负面引用的时间性质；建立反映负面引用的符号科学引文网络，运用符号网络分析的指标和方法，从复杂网络角度对负面引用的拓扑性质进行探索；此外，运用不同的排名算法对论文影响力进行计算，并与普通引文网络影响力排名结果对比，计算并对比各排名结果之间的相关性，探索负面引用对于论文排名的影响。
首先，对于负面引用的文本特征进行分析。对已经人工标注好情感倾向的AAN引文语料数据与其他三个网站评论语料数据，采取同样的情感得分计算方式，分析得分结果的差异。通过对比，验证了相较于日常网站评论中的负面情绪，科学引文中的负面表达在词性上会更含蓄。
其次，分析负面引用的时间特性。论文收到的负面评价出现在初次被引前期的概率较大；对比被引前后论文收到的引用的变化，负面引用则在一定程度上会降低论文获得的关注度；整体而言，获得明显负面情感的论文占少数；随着时间的推移，文章更多地收到一些相对较为客观的中性评价。
此外，构建符号引文网络分析负面引用的拓扑性质。基于ACL ARC数据集建立符号网络，并计算基本统计指标，同时通过与其交叉换边后的随机网络进行对比分析，发现负面引用的相关拓扑结构性质：度分布多为厚尾分布；负面引用中高度值的节点更倾向于和低度值节点相连，连通性远低于正面引用；与随机网络相比，原网络整体平衡性更高，而负面引用网络的集聚程度和网络连通性更高。
最后，探索负面引用对于论文排名的影响。本文分别从传统文献计量和复杂网络结构两个角度，基于是否考虑负面引用因素，共采用四种论文影响力计算指标，对ACL ARC论文数据进行排名并对比分析结果：从论文排名的差异来看，在复杂网络结构视角下考虑负面引用时，它对于排名过高或过低的论文影响更明显，对于整体排名的影响相对较小；相比于整体排名，高排名论文和排名靠后论文的相关度更小。

As scholars explore the laws of scientific development, the study of scientific systems is gradually receiving more attention from the academic community. Scientific citations represent the citation connections between papers, reflect the processes of the propagation scientific achievements, and can also assist in measuring the influence of papers. At present, most researches on citations focus on statistical indicators in macro level, and often assume these citations are positive, but in reality, there are also some citation relationships with negative tendencies. These negative citations also play an important role in measuring the impact of scientific papers. Exploring the characteristics of interactions between papers through negative citation relationships is of great significance for the study of scientific systems.
This article analyzes AAN citation corpus and ACL ARC citation relationship data, uses sentiment recognition algorithms to calculate sentiment scores, compares and analyzes the differences in negative expressions between scientific citations and daily comments. Using time data of citations, we calculate the changes of relevant statistical indicators over time, and explore the temporal characteristics of negative citations.We establish a signed citation network which can reflect negative citations, and use symbolic network analysis indicators and methods to explore the characteristics of negative citations. In addition, we calculate the influence value of the paper, compare it with the ranking results of ordinary citation network influence value, and analyze the impact of negative citations on the ranking of the paper.
Firstly, analyze the textual features of negative citations. Compare the results using the same emotional score calculation method between the AAN citation corpus data and the comment corpus data of three other websites. It has been verified that compared to negative emotions in daily website comments, negative expressions in scientific citations are more implicit in terms of part of speech.
Secondly, analyze the temporal characteristics of negative citations.The probability of the negative citations received by a paper appearing in the early stage of publication is relatively high.Comparing the changes in citations received by the paper before and after being cited, negative citations will reduce the attention received by the paper. A small percentage of papers receive significantly negative emotions. Over time, the articles have received more objective and neutral evaluations.
In addition, constructing a signed citation network to analyze the topological properties of negative citations. We establish a signed network based on the ACL ARC data and calculate basic statistical characteristics. By comparing and analyzing the random network after swapping links, it is found that the topological characteristics of negative citations are: degree distribution is mostly a thick tailed distribution. Nodes with high degrees in negative references tend to be more connected to nodes with low degrees, and their connectivity is much lower than that in positive citations. Compared with random networks, negative citation networks have a more pronounced degree of aggregation and a higher overall network balance.
Finally, explore the impact of negative citations on paper ranking. From the perspectives of traditional bibliometrics and complex networks, based on whether negative citations are considered or not, different paper influence calculation indicators are used to rank ACL ARC data and compare the analysis results. From the difference in paper rankings, negative citations have a more significant impact on the overall ranking, but have a smaller impact on papers with high or low rankings. Compared to the overall ranking, papers with high rankings have lower relevance, while lower ranked papers have higher overall relevance than high ranked papers. In the early stages of publication, the probability of receiving great evaluations is high for a paper. Overall, a small number of papers receive significantly negative emotions, while most papers receive neutral evaluations.