Task-Oriented Semantic Communication with Large Language Model Enabled Knowledge Base
Wuxia Hu, Caili Guo, Zhiyu Zhu, Yang Yang, Chunyan Feng · 2025
Large Language Models (LLMs) have the potential to greatly enhance the performance of task-oriented semantic communication (TOSC) through their extensive knowledge. However, hallucinations from LLMs may cause semantic noise, thus degrading task performance. To address this issue, we propose a TOSC scheme with an LLM-enabled knowledge base (TOSCLKB). In the considered system, the transmitter transmits the features of the source over a wireless channel to accomplish downstream tasks at the receiver. Meanwhile, the transmitter utilizes the LLM-enabled Knowledge Base (KB) to provide additional data for downstream tasks. To effectively exploit the extensive knowledge of LLMs while mitigate its hallucination, a cross-domain fusion codec framework with a hallucination filtering phase and a cross-domain fusion phase is proposed. In particular, the first phase filters out data irrelevant to the source generated by the LLM-enabled KB based on semantic similarity. Then, a cross-domain fusion phase is proposed which fuses source data with LLM-generated data based on their semantic importance, thereby enhancing task performance. Experiment results on the text-based person retrieval task demonstrate that the proposed TOSC-LKB can achieve up to 26.7 % and 7.1 % performance gains without introducing additional time overhead compared to DeepSC and TOSC-LKB without LLMs.