Labeling short, unstructured texts is generally performed by sequentially identifying codes and assigning them to segments of text based on viewing a small sample of data. In this greedy approach, coders risk overlooking important code ideas and must perform the tedious task of iteratively revising the initial code set, and sometimes response code assignments, as new themes emerge. To address this, we propose CodeML, a machine learning-assisted (ML) coding interface that identifies multiple ideas in a response, which are displayed to support interactive data exploration, code identification, and refinement of snippet code assignments. By surfacing themes and snippets early, coders can consider a broader range of potential codes to reduce chances of omitting codes that surface later. A comparative study against search-style coding shows the potential for CodeML to facilitate initial exploration and discovery of finer-grained code sets while not adding significant cognitive load to organize codes and underlying text snippets. READ MORE