As advanced technologies, such as data mining become part of the everyday workflow of document reviews in litigations, keyword-search still appears to serve as a cornerstone approach in responsive or privilege review. Keywords are conceptually easy to understand and help culling documents at the early stages of the review. But developing proper keywords to minimize the risk of under/over-inclusiveness can lead to complex strategies. To cope with the burden of designing search terms, we propose to use word embedding techniques in a dynamic manner. This paper describes a system leveraging semantic models in a smart review environment in order to support knowledge workers in eDiscovery.