Text Semantic Communication Systems with Sentence-Level Semantic Fidelity

Semantic communication systems have been proposed for efficient text transmissions in recent years. However, most existing methods mainly focus on word-level recovery, which fails to reflect the semantic fidelity of the model. By leveraging the sentence-level semantic information, a semantic communication system architecture is proposed within the framework of deep learning based joint source-channel coding. Then, in order to evaluate the semantic fidelity of the system, a new metric of Semantic Similarity is proposed, which is more sensitive to semantic differences as compared to existing evaluation metrics, e.g., bilingual evaluation understudy. For guaranteeing the consistency between model training and performance evaluation, the proposed new metric is then incorporated into the objective function, based on which end-to-end performance optimization is performed. Extensive simulation results on European Parliament dataset demonstrate the effectiveness and necessity of the proposed method. Compared with the state-of-the-art, e.g., DeepSC, an improvement of up to 10% is achieved in terms of Semantic Similarity. Furthermore, significant performance gains are achieved by the proposed method in both regimes of low and medium signal-to-noise ratio, as compared to traditional separate source-channel coding methods.