Semantic Error Correction for Short Block Codes Over Noisy Channels

A newly developed receiver framework improves the transmission of natural language sentences across noisy wireless channels by utilizing several short block codes. Sentences are first ASCII-encoded and segmented, with each segment independently encoded using a short block code prior to being sent over an AWGN channel. Upon reception, segments are decoded simultaneously, and a semantic error correction (SEC) model is employed to reconstruct any corrupted segments by leveraging language model context. Additionally, the framework features semantic list decoding (SLD), which produces multiple possible reconstructions and identifies the optimal one based on weighted Hamming distance, along with a semantic confidence-guided HARQ (SHARQ) mechanism that substitutes CRC-based error detection with a confidence score for selective retransmission. All components are developed and trained with bidirectional and auto-regressive transformers (BART), and simulation results are included.