基于pu-learning的同行评议文本情感分析

最近几年逐渐出现了对同行评议文本情感分析的研究，包括通过同行评议文本预测审稿人的推荐状态的任务。现有模型融入了论文本身或摘要信息，采用神经网络学习论文或摘要的高层表示，结合同行评议文本预测审稿人的推荐状态，这使得模型变得非常复杂的同时结果却没有实质性的提高。为此，提出了OSA机制来提高情感分析模型中对观点句的关注度。具体来说，采用pu-learning从同行评议文本的前N个句子中学习观点句的特征，使每一个句子都得到一个观点句权重，将其应用于情感分析模型的倒数第二层，由此得到最终的预测结果。在ICLR2017—2018数据集上使用不同的情感分析模型对OSA进行了评估，实验结果验证了OSA的高效性，并在两个数据集上取得了优异的性能。

Sentiment Analysis of Peer Review Texts Based on Pu-Learning

There have been some researches on the sentiment analysis of peer review text, including the task of predicting the overall recommendation through a peer review text written by reviewer for a submission. Existing works integrate the embedding of the paper or abstract, utilizing neural network to learn the high-level representation of paper or abstract and review text to predict reviewer’s overall recommendation, which make the algorithm very complicated but the effect is not substantially improved. To solve this issue, a mechanism called OSA（opinionated sentence attention） is proposed to make opinionated sentences get more attention in sentiment analysis model. Specifically, this paper employs a positive-unlabeled learning method to learn opinionated sentence features form Top-N sentences of peer review texts so that every sentence of all review texts gets a opinionated weight, then these weights are dotted with penultimate layer of neural network to get the final prediction. OSA is evaluated with different neural networks on ICLR 2017—2018 datasets, experimental results verify that OSA is of high efficiency and achieves outstanding performance on two datasets.