Unlike traditional image retrieval methods that allow users to express their search intent solely through text or image queries, Composed Image Retrieval (CIR) enables users to take advantage of multimodal queries, where the reference image is combined with a modification text to retrieve the target image. In recent years, mainstream research on Compositional Image Retrieval (CIR) typically employs large pre-trained vision-language models as feature extraction backbones, followed by nonlinear feature-level multimodal query fusion for target image retrieval. Furthermore, a solution involving raw-data level multimodal feature fusion has been proposed to prevent fused features from deviating from the original embedding space, which could potentially harm retrieval performance. While this approach does alleviate the feature bias issue to some extent, we argue that directly handling the fused raw data may lead to problems such as feature loss and multimodal matching issues. To address this, we propose a pre-training phase for the text encoder, which learns a semantic modification strategy by combining descriptions of candidate images, modification text, and target text. For each query, we pre-select the top-K most similar target images as global soft labels, and the target image probability distribution within the batch serves as local soft labels. This approach guides the similarity probability distribution between queries and targets, pushing refined query features and false negative sample features closer together in the latent space, thereby alleviating multimodal matching loss. We validated the effectiveness of our improvements through experiments on three public datasets.