基于短文本情感分析的敏感信息识别
Sensitive Information Recognition Based on Short Text Sentiment Analysis

针对现有的敏感信息识别是基于敏感关键词匹配方式判断的,准确度不是很高,且具有较高的误检率等问题,提出了敏感关键词与情感极性协同分析的敏感信息识别方法。在真实数据集上,利用监督学习的方式对微博的情感极性进行了度量,得到具体的情感极度,并将文本分为正情感极性和负情感极性两类。通过定义色情、暴力、违禁、邪教、反动等5大类2 639个敏感关键词和在数据集里面所呈现的Zipf分布特性,发现含有负情感极性的微博内容具有较高的敏感性,由此深入研究了敏感关键词对情感极性的动力因素,构建了含有情感极性因素的敏感度模型,提出了敏感信息的判别方法,敏感信息检测的准确率由传统方法的31.25％提高到了58.75％,召回率则由95％提升到96％,综合指标F值从47.0％提升到了72.3％。
The existing sensitive information recognition is based on the sensitive keyword matching method, so the accuracy is low and the miss rate is high. We presented a collaborative method by using the sensitive keywords and sentiment polarities to identify the sensitive information. In the real dataset, we used the supervised way to measure the sentiment polarities of the blogs, and divided the blogs into two categories, namely the blogs are with positive or negative sentiment polarities. Five kinds of 2 639 sensitive keywords, including pornography, violence, illegality, cult and reactionary, were defined, and it was found that according to the Zipf distribution of these words in the dataset, the contents of blogs with negative sentiment polarities exhibited high sensitivities. Then we studied the contribution of the sensitive keywords to the sentiment polarity, and constructed the model of sensitivity degree that contains the sentiment polarity factor. Based on this, we proposed a new way to identify the sensitive information, which makes the accuracy and miss rate improved from 31.25％ to 58.75％ and from 95％ to 96％, respectively, and the F？？measure was improved from 47.0％to 72.3％