Engineering Semantic Communication: A Survey

Abstract

As the global demand for data has continued to rise exponentially, some have begun turning to the idea of semantic communication as a means of efficiently meeting this demand. Pushing beyond the boundaries of conventional communication systems, semantic communication focuses on the accurate recovery of the meaning conveyed from source to receiver, as opposed to the accurate recovery of transmitted symbols. In this survey, we aim to provide a comprehensive view of the history and current state of semantic communication and the techniques for engineering this higher level of communication. A survey of the current literature reveals four broad approaches to engineering semantic communication. We term the earliest of these approaches classical semantic information, which seeks to extend informationtheoretic results to include semantic information. A second approach makes use of knowledge graphs to achieve semantic communication, and a third utilizes the power of modern deep learning techniques to facilitate this communication. The fourth approach focuses on the significance of information, rather than its meaning, to achieve efficient, goal-oriented communication. We discuss each of these four approaches and their corresponding studies in detail, and provide some challenges and opportunities that pertain to each approach. Finally, we introduce a novel approach to semantic communication, which we term contextbased semantic communication. Inspired by the way in which humans naturally communicate with one another, this context-based approach provides a general, optimization-based design framework for semantic communication systems. Together, this survey provides a useful guide for the design and implementation of semantic communication systems.

Deep Learning-Empowered Semantic Communication Systems With a Shared Knowledge Base

“…where 𝜆 is a positive number to balance the quantization relation between the first term and the second term. According to (9), the term (1−𝐷 (s, k * s )) stands for the sentence similarity, which can be evaluated by using BLEU (4) or sentence similarity score (6). When s = k s , the semantic distance between the message s and the knowledge k s approaches zero, resulting in 𝐷 (s, k s ) = 0 and consequently 𝐼 𝑠 (s) = 𝐼 (k s ).…”

“…Problem 1: In this paper, the objective of knowledge base construction is to minimize the expectation of the distance between the message s and its most similar knowledge k * s identified by (9) among the knowledge base 𝔎, i.e.,…”

“…where 𝐻 𝑠 (K/S) denotes the semantic entropy of K conditioned to S and 𝐻 𝑠 (S/K) denotes the semantic entropy of S conditioned to K. According to (9), for each message, there is a unique piece of knowledge in the knowledge base corresponding to it, which means 𝐻 𝑠 (K/S) = 0. Besides, we use residual information r to represent the obtained information that contained in s but not in k. Normally, the semantic entropy of residual information, 𝐻 𝑠 (R) is not greater than 𝐻 𝑠 (S|K), due to the fact that only a finite number of complex symbols are used to transmit information.…”

“…These issues motivate a paradigm shift from conventional communications to semantic communications [8], [9], which extract semantic meanings behind bits for transmission and recover the semantic information at the receiver through minimizing the semantic error. There are two main advantages of semantic communications over conventional communications.…”

“…Without loss of generality, most of the existing semantic systems can be divided into two categories according to different communication modalities, i.e., human-to-human (H2H) and human-to-machine (H2M) communications [15], [16]. To be specific, H2H communications aim to deliver meanings accurately over a channel for message exchanges between two human beings [9]. While in H2M communications, the message sent by a human being should be correctly interpreted by a machine toward performing specific tasks such as speech to text transmission [17], image classification [18], spectrum sensing [19] and visual question answering [20].…”