Text-based person search aims to retrieve the pedestrian images that best match a given text query. Existingmethods utilize class-id information to get discriminativeand identity-preserving features. However, it is not wellexplored whether it is beneficial to explicitly ensure that thesemantics of the data are retained. In the proposed work, weaim to create semantics-preserving embeddings through anadditional task of attribute prediction. Since attribute annotation is typically unavailable in text-based person search,we first mine them from the text corpus. These attributes arethen used as a means to bridge the modality gap between theimage-text inputs, as well as to improve the representationlearning. In summary, we propose an approach for textbased person search by learning an attribute-driven spacealong with a class-information driven space, and utilizeboth for obtaining the retrieval results. Our experiments onbenchmark dataset, CUHK-PEDES, show that learning theattribute-space not only helps in improving performance,giving us state-of-the-art Rank-1 accuracy of 56.68%, butalso yields humanly-interpretable features.