The term ESG refers to the distinct evaluation of sustainability practices from three aspects. These are namely environmental, social and governance. As the incorporation of these factors into investments has gained popularity, many asset managers have begun to measure their ESG risks. The assessment of these risks are conducted by various rating agencies based on ESG disclosures of companies. The conflicting views that have accumulated in academia regarding the reliability of ESG ratings mostly originate from the fact that ratings can be interpreted in different ways. The work presented in this thesis aims to solve the problem of the inherent vagueness involved with evaluating and quantifying the actions of a company with respect to the three categories of ESG. We focus on the social dimension of ESG, to create a rating method that uses the likelihood of controversy occurrence as a basis. To produce this rating methodology, we make use of highly capable machine learning algorithms to detect patterns that lead to controversy, using a wide indicator set. The most important among these indicators are identified to be used practically in investment decisions.