Text-based person re-identification aims to retrieve the matched persons’ images given a textual query. Recently, the literature has widely-adopted the powerful CLIP to extract fine-grained visual/textual features for cross-modal alignment. Yet, they generally overlook the seemingly similar characteristics in person re-identification, namely, the visuals and descriptions of two persons may share many intersections (e.g., grey jacket, walking), potentially leading to false matches. This work presents a novel framework by carefully elaborating on the identifiable CLIP tokens for the explicit attention of matching. The token informativeness of local features is measured by the visual/textual entropy related to person clothing, posture, hairstyle, etc. Additionally, a triplet ranking loss function is incorporated to expand both intra-modal and inter-modal distances among different persons/descriptions. Experimental results on three datasets demonstrate the superiority of our proposal against prior baselines.