GenSC: Generative Semantic Communication Systems Using BART-Like Model

Abstract

The current mindset of semantic communications focuses on how to have the sentence received exactly. However, as long as the received sentence and the original sentence are perceived the same or similarly, it can be reviewed as a successful semantic-level transmission. Hence, we design a new architecture for the semantic communication system based on BART-like model (Lewis et al., 2019), called GenSC. The proposed GenSC further takes the token-level correlation between consecutive tokens into account during the semantic encoding and this bidirectional correlation helps correct or fill in a “semantically similar token” at the semantic decoder when a token is missing or corrupted during transmission. The simulation shows that compared to conventional approaches such as Xie et al. (2021) and Liu et al. (2022), GenSC can improve the bilingual evaluation understudy (BLEU) and semantic similarity (SS) scores at low SNR regions a lot and enjoy higher BLEU and SS scores at high SNR regions. When SNR is 0dB, GenSC outperforms (Xie et al., 2021) and (Liu et al., 2022) by around 30% (resp. 84%) and 18% (resp. 55%) in terms of BLEU (resp. SS), respectively.