Utilizing Lexicon-enhanced Approach to Sensitive Information Identification

Large-scale sensitive information leakage incidents have occurred frequently, causing huge impacts and losses to individuals, enterprises, and society. Most sensitive information exists in unstructured data, making it challenging for people to identify when it is leaked, an important cause of information leakage. Therefore, sensitive information identification from unstructured data has received extensive attention. In addition, the smallest unit of Chinese is a character, so its lexical boundary is flexible, which makes it very difficult to identify sensitive information in Chinese. It is worth mentioning that there are no publicly available datasets in this field of sensitive information identification due to the sensitivity. To address the above challenges, we first create the SPIDC (Sensitive Personal Information Dataset in Chinese) and release it as a public resource for related research. Second, we apply the existing sensitive information identification methods on the English datasets to the Chinese datasets. In addition, to solve the problem of uncertainty and ambiguity of Chinese vocabulary boundary, we apply three lexicon-enhanced technologies from NER (Named Entity Recognition) to the Chinese sensitive information identification for the first time. Experimental results on the SPIDC show that the lexicon-enhanced approach has better performance than other methods.