Nowadays, the publication of textual documents provides critical benefits to scientific research and business scenarios where information analysis plays an essential role. Nevertheless, the possible existence of identifying or confidential data in this kind of documents motivates the use of measures to sanitize sensitive information before being published, while keeping the innocuous data unmodified. Several automatic sanitization mechanisms can be found in the literature; however, most of them evaluate the sensitivity of the textual terms considering them as independent variables. At the same time, some authors have shown that there are important information disclosure risks inherent to the existence of relationships between sanitized and non-sanitized terms. Therefore, neglecting term relationships in document sanitization represents a serious privacy threat. In this paper, we present a general-purpose method to automatically detect semantically related terms that may enable disclosure of sensitive data. The foundations of Information Theory and a corpus as large as the Web are used to assess the degree relationship between textual terms according to the amount of information they provide from each other. Preliminary evaluation results show that our proposal significantly improves the detection recall of current sanitization schemes, which reduces the disclosure risk.