The peer review of academic papers can directly reflect the subjective evaluation of reviewers on the papers, and the extraction of sentiment information from the review text is beneficial to mining rich information of reviewers’evaluation on each dimension of the papers. The existing sentiment analysis task could only extract the single review dimension and sentiment of experts. A fine-grained sentiment analysis model for peer review is proposed based on multi-task learning. The model is equipped with the ability to accomplish both attribute word extraction and fine-grained sentiment analysis tasks by adding the BiLSTM-CRF module to the BERT-LCF model in a multi-task learning framework. Compared with the traditional single-task fine-grained sentiment analysis model based on the Pipeline model, the proposed model can complete the review attribute extraction and sentiment analysis tasks simultaneously while ensuring the accuracy of the model. In the two tasks, F1-score of the proposed model reaches 89.01% and 90.71%, respectively. In addition, the introduction of BiLSTM-CRF module has a certain enhancement effect on the review text attribute word extraction task in a multi-task scenario, as demonstrated by comparison experiments.
常见的细粒度情感分析模型主要采用深度学习方法,如TD-LSTM[4](Target-dependent long short term memory)、 ATAE-LSTM[5](Attention-based LSTM with aspect embedding)和IAN[6](Interactive attention networks)。这些模型都利用了长短期记忆网络(Long short term memory)和注意力机制(Attention),在文本序列建模的基础上,通过注意力机制将情感极性的预测聚焦于与之对应的属性词上。随着大规模预训练模型的兴起,在细粒度情感分析领域,以BERT为输入层的模型取得较高的性能提升。基于BERT开发的BERT-PT[7](BERT post-training)、 BERT-SPC[8]和BERT-LCF[9](BERT local context focus),都是在传统的BERT输入序列末尾通过拼接属性词的方式,使得BERT模型更关注属性词序列的相关信息,以此提高该属性情感倾向的预测效果。然而,由于此类数据的标注是一个既费时又费力的工作,大多数研究往往不会提供语句里属性词的标记。因此,现有的细粒度情感分析任务很少会去考虑属性词的抽取任务,造成其在专有领域的应用性研究较少。
BLBC模型任务为给定一个包含属性的同行评审句子序列 S ={w1,…,wm,wm+1,…,wm+k,…,wn },其中,为句子中的单词为属性词序列,n为句子的长度,输出为{属性词,情感极性},模型结构如图 2 所示。
2.1 BERT共享输入层
BERT(Bidirectional encoder representation from transformers)是一种基于大规模预训练的语言表征模型,能表征文本丰富的语义信息[17]。输入层采用基于BERT共享层的词嵌入方式,使用BERT-BASE(如图3)和预训练模型BERT-SPC(如图4),分别构建属性抽取输入表征模型和情感分析输入表征模型,用以提取全局和局部上下文信息。这两个模型的输入都由词向量、块向量和位置向量组成,区别在于输入形式不同。在BERT输入层中,以两个独立的BERT进行编码。全局输入序列为,局部输入序列为,其中,为输入句子序列,为句子中的属性词序列。
BiLSTM-CRF为属性词序列标记模型,如图 5 所示。在对输入特征进行双向LSTM编码后,条件随机场CRF(Conditional random fields)可以学习前后文属性词标签之间的约束关系,以降低属性词提取的错误率,得到具有最大概率的合理序列[19]。在CRF中,设全局特征经过BiLSTM(Bidirectional long short-term memory)编码后为,其对应的属性词标签预测结果的得分计算方式为
BENOSD J, BASHARIE, CHAVESJ M,et al. The ups and downs of peer review[J].Advances in Physiology Education,2007,31(2):145-152.
[2]
BORNMANNL, DANIELH D.Reliability of reviewers’ ratings when using public peer review:a case study[J].Learned Publishing,2010,23(2):124-131.
[3]
RAGONEA, MIRYLENKAK, CASATIF,et al.On peer review in computer science:Analysis of its effectiveness and suggestions for improvement[J].Scientometrics,2013,97(2):317-356.
[4]
TANGD, QINB, FENGX,et al.Effective LSTMs for target-dependent sentiment classification[J].2015,arXiv:
[5]
WANGY, HUANGM, ZHUX,et al.Attention-based LSTM for aspect-level sentiment classification[C]//2016 Conference on Empirical Methods in Natural Language Processing.2016:606-615.
[6]
MAD, LIS, ZHANGX,et al.Interactive attention networks for aspect-level sentiment classification[C]//26th International Joint Conference on Artificial Intelligence.2017:4068-4074.
[7]
XUH, LIUB, SHUL,et al.BERT post-training for review reading comprehension and aspect-based sentiment analysis[C]//NAACL-HLT.2019:2324-2335.
[8]
SONGY, WANGJ, JIANGT,et al.Attentional encoder network for targeted sentiment classification[J].arXiv:2019.
[9]
ZENGB, YANGH, XUR,et al.LCF:A local context focus mechanism for aspect-based sentiment classification[J].Applied Sciences,2019,9(16):3389.
[10]
WANGK, WANX.Sentiment analysis of peer review texts for scholarly papers[C]//The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval.2018:175-184.
[11]
GHOSALT, VERMAR, EKBALA,et al.Deepsentipeer:Harnessing sentiment in review texts to recommend peer review decisions[C]//The 57th Annual Meeting of the Association for Computational Linguistics.2019:1120-1130.
LINYuan, WANGKaiqiao, YANGLiang,et al.Sentiment analysis of peer review texts based on pu-Learning[J].Computer Engineering and Applications,2023,59(3):143-149. (in Chinese)
[14]
YUANW, LIUP, NEUBIGG.Can we automate scientific reviewing?[J].Journal of Artificial Intelligence Research,2022,75:171-212.
[15]
CHAKRABORTYS, GOYALP, MUKHERJEEA.Aspect-based sentiment analysis of scientific reviews[C]//ACM/IEEE Joint Conference on Digital Libraries in 2020.2020:207-216.
DINGShengchun, FANGZhen, WANGNan.Business domain named entity recognition based on Bi-LSTM-CRF[J].Journal of Modern Information, 2020,40 (3):103-110. (in Chinese)
[23]
YANGH, ZENGB, YANGJ H,et al.A multi-task learning model for chinese-oriented aspect polarity classification and aspect term extraction[J].Neurocomputing,2021,419:344-356.