PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Communications for Future Internet: Fundamentals, Applications, and Challenges

Wanting Yang, Hongyang Du, Ziqin Liew, Wei Yang Bryan Lim, Zehui Xiong, Dusit Niyato, Xuefen Chi, Xuemin Sherman Shen, Chunyan Miao

arXiv:2207.00427v2cs.NIeess.SP

TL;DR

SemCom-enabled 6G lacked a comprehensive survey of its developments, challenges, and future trends. This paper surveys SemCom for 6G by tracing its evolution, classifying modern approaches, organizing system design dimensions, and reviewing applications, challenges, architecture, and future directions.

  • Problem

    SemCom-enabled 6G lacked a comprehensive survey covering its developments, challenges, and future trends.

  • Method

    The paper surveys SemCom for 6G, tracing its evolution, classifying three types, and organizing communication-system design around SI extraction, SI transmission, and SI metrics.

  • Results

    The survey presents SemCom’s applications, SemCom-empowered network architecture, state-of-the-art techniques, challenges, and future research directions for 6G.

  • Takeaways & Limitations

    The paper provides a guide for researchers and practitioners incorporating SemCom concepts into future communication architectures.

  • Takeaways & Limitations

    The survey notes that implicit SemCom remains difficult because receivers must infer meanings from transmitters’ context and background.

Abstract

from arXiv · show

With the increasing demand for intelligent services, the sixth-generation (6G) wireless networks will shift from a traditional architecture that focuses solely on high transmission rate to a new architecture that is based on the intelligent connection of everything. Semantic communication (SemCom), a revolutionary architecture that integrates user as well as application requirements and meaning of information into the data processing and transmission, is predicted to become a new core paradigm in 6G. While SemCom is expected to progress beyond the classical Shannon paradigm, several obstacles need to be overcome on the way to a SemCom-enabled smart wireless Internet. In this paper, we first highlight the motivations and compelling reasons of SemCom in 6G. Then, we outline the major 6G visions and key enabler techniques which lay the foundation of SemCom. Meanwhile, we highlight some benefits of SemCom-empowered 6G and present a SemCom-native 6G network architecture. Next, we show the evolution of SemCom from its introduction to classical SemCom related theory and modern AI-enabled SemCom. Following that, focusing on modern SemCom, we classify SemCom into three categories, i.e., semantic-oriented communication, goal-oriented communication, and semantic-aware communication, and introduce three types of semantic metrics. We then discuss the applications, the challenges and technologies related to semantics and communication. Finally, we introduce future research opportunities. In a nutshell, this paper investigates the fundamentals of SemCom, its applications in 6G networks, and the existing challenges and open issues for further direction.

I. INTRODUCTION

The introduction motivates SemCom as a response to 6G’s data-, resource-, and service demands, and positions the survey as a comprehensive guide to SemCom-enabled 6G networks. It outlines the survey’s categories, design dimensions, applications, challenges, and future directions.

  • I. INTRODUCTION: Content-centric, data-driven architectures can hinder high-quality user services as emerging 6G applications become human-centric, data-intensive, and resource-intensive.
  • I. INTRODUCTION: 6G applications require massive data transmission, fast responses, reliable information interaction, and continuous information updates.
  • I. INTRODUCTION: SemCom uses message meaning and transmits information relevant to receivers or communication tasks after AI-based data preprocessing.
  • I. INTRODUCTION: SemCom and 6G are mutually reinforcing: distributed computation and ubiquitous AI support deployment, while SemCom addresses traditional communication constraints.
  • I. INTRODUCTION: The survey classifies SemCom into semantic-oriented, goal-oriented, and semantic-aware communication, distinguishing connection-oriented and task-oriented communication.
  • I. INTRODUCTION: The survey organizes SemCom system design around SI extraction, SI transmission, and SI metrics, then reviews techniques, limitations, applications, challenges, and future research.

II. FUNDAMENTALS OF SEMCOM

This section traces SemCom from semiotic and Shannon-based foundations toward semantic and effectiveness levels, then organizes modern SemCom around extraction methods and semantic metrics. It also highlights that semantic performance evaluation remains less developed than extraction and transmission.

  • II. FUNDAMENTALS OF SEMCOM: Semiotics distinguishes syntactics, semantics, and pragmatics, corresponding respectively to formal signs, meaning, and utility relative to users.
  • II. FUNDAMENTALS OF SEMCOM: Shannon’s communication theory focuses on the technical level, motivating extensions toward semantic meaning and communication effectiveness.
  • II. FUNDAMENTALS OF SEMCOM: Classical semantic information theory models sentence information using a measurement function and entropy-based semantic information calculation.
  • II. FUNDAMENTALS OF SEMCOM: The survey reviews four semantic extraction approaches: DL-based, RL-based, KB-assisted, and semantic-native extraction.
  • II. FUNDAMENTALS OF SEMCOM: Semantic metrics are discussed for text, visual, and audio data, alongside error-, AoI-, VoI-, and combined metric categories.

2) Strongly semantic information theory:

Strongly semantic information theory extends semantic information measurement by incorporating truth values and discrepancy from the actual situation. Its proposed metrics impose conditions for truth, tautology, contradiction, and contingent statements, but their scope remains limited.

  • 2) Strongly semantic information theory:: Strongly semantic information theory incorporates truth values, extending weakly semantic information theory’s treatment of semantic information.
  • 2) Strongly semantic information theory:: The discrepancy function f(s) assigns 0 to the most accurate true statement, 1 to tautologies, and −1 to contradictions.
  • 2) Strongly semantic information theory:: Contingently false statements satisfy −1 < f(s) < 0, whereas contingently true statements also true in other situations satisfy 0 < f(s) < 1.
  • 2) Strongly semantic information theory:: The approach quantitatively analyzes complete classes of propositions in logical space but does not provide rigorous metrics beyond that scope.

3) Semantic communication theory:

Semantic communication theory evolved from analytical models of semantic information and shared knowledge toward practical system models that target meaning and communication goals. The survey organizes modern SemCom into semantic-oriented, goal-oriented, and semantic-aware communication while distinguishing semantic processing from classical communication.

  • Early SemCom theory modeled sources and destinations using world models, background knowledge, inference, and message interpretation to support semantic-level communication.
  • Semantic channel capacity depends on mutual information, semantic ambiguity from encoding, and the logical information of received messages; mismatched knowledge or inference generates semantic noise.
  • Initial theoretical frameworks remained limited to simple logic-language or server-printer scenarios and quantified information with Shannon entropy rather than meaning-sensitive measures.
  • Goal-oriented communication: Goal-oriented communication extends semantic models by distinguishing metagoals, which capture agents’ intents, from syntactic goals, which capture observable effects.
  • SemCom categories: The survey classifies SemCom into semantic-oriented, goal-oriented, and semantic-aware communication, corresponding to different levels and roles of semantics in communication.
  • Semantic-oriented communication: Semantic-oriented communication prioritizes semantic-content accuracy over average source-data information and adds semantic representation before encoding to retain core information while filtering redundancy.

2) Goal-oriented communication:

Goal-oriented communication incorporates the communication task into semantic extraction and produces actions rather than merely recovered meaning. It emphasizes task effectiveness under limited resources, while requiring consistent local knowledge and goals.

  • Goal-oriented communication: Goal-oriented communication captures pragmatic information relevant to a specific communication goal, unlike semantic-oriented communication’s focus on semantic information.Pragmatic information is treated as task-relevant semantic information conveyed through syntactic information.
  • Goal-oriented communication: Its semantic extraction uses the communication goal and shared local knowledge to filter irrelevant information when goals change.The goal must play an important role in semantic extraction and be available to communication parties.
  • Goal-oriented communication: The system output is a direct action, such as acceleration, braking, steering, or flashing headlights, rather than recovered meaning.These actions respond to pedestrians, roadblocks, and traffic-signal changes.
  • Goal-oriented communication: Goal-oriented communication targets effective task completion under limited network resources rather than semantic-information accuracy.This distinguishes its effectiveness-level objective from semantic-oriented communication’s semantic-level objective.
  • Goal-oriented communication: Consistent local knowledge and communication goals are required across parties; otherwise semantic noise may cause task failure.The section also contrasts goal-oriented systems with semantic-aware communication, which may lack explicit transceivers and a general system model.
  • Semantic extraction technologies: Semantic extraction in SemCom converts understanding-before-transmission into a strategy for addressing bandwidth bottlenecks.The survey reviews extraction methods across semantic-oriented, goal-oriented, and semantic-aware communication.

2) SE for text data:

Semantic extraction for text data evolved from embedding- and recurrent-network methods toward attention-based architectures. These approaches improve text recovery under changing channels, while related audio methods extend deep-learning semantic extraction beyond text.

  • SE for text data: The pioneering text SemCom system uses GloVe word embeddings with LSTM encoder-decoder networks for transmission over an erasure channel.Beam search selects likely word sequences during decoding.
  • SE for text data: LSTM-based semantic extraction achieves the lowest word error rate at a given coding length with a high bit-drop rate compared with Gzip and Huffman.The passage attributes the result to the effectiveness of extracted information.
  • SE for text data: Transformers use multi-head attention to extract semantic information and syntax from whole sentences in parallel.They are presented as an alternative to recurrent architectures such as LSTM.
  • SE for text data: A Universal Transformer adds adaptive circulation to handle semantic ambiguity and noise more flexibly than a fixed attention structure.The approach is motivated by the differing importance of words and phrases during sentence processing.
  • SE for text data: Across the full SNR region, the UT-based SemCom algorithm consistently achieves higher BLEU scores than the Transformer-based algorithm, while both outperform cascaded conventional schemes under varying channels.The conventional schemes improve substantially only above 15 dB SNR.
  • SE for audio data: Audio semantic extraction extends deep-learning designs using Wav2Vec, while SE-ResNet assigns higher weights to essential information during training.The audio encoder uses feature extraction and feature aggregation modules.

4) SE for multimodel data:

The survey extends semantic extraction to multimodal, reinforcement-learning, knowledge-base, and emergent-communication settings. These methods target task-relevant information and contextual adaptation, but face reward-design, training-data, and system-model limitations.

  • SE for multimodal data: A multimodal VQA SemCom system jointly transmits images and questions, producing answers at the receiver through end-to-end processing.It achieves significantly higher answer accuracy than recovering image and text before conventional multimodal processing, but assumes perfect channel state information.
  • RL-based SE: Reinforcement learning addresses user-defined, task-specific, and non-differentiable semantic metrics by treating token generation as sequential actions.The encoder-decoder interacts with sentences as an environment, and whole-sentence semantic metrics serve as long-term returns.
  • RL-based SE: Immediate reward design is difficult because sentence-level rewards are unavailable until decoding ends, while alternative estimators can consume substantial resources or diverge.Self-critical sequence training is introduced to normalize long-term rewards using test-time inference.
  • Knowledge-base-assisted SE: For multitask communication, semantic information differs by goal, so a shared knowledge base can associate communication goals with compatible semantic-information combinations.The knowledge base stores semantic information, task goals, and jointly understandable reasoning rules.
  • Knowledge-base-assisted SE: Knowledge-base-assisted SemCom with CR3 of 98% achieves more than 40% classification-accuracy gains over conventional communication at 10 dB.The passage notes remaining room to improve neural-network structure and loss-function optimization.
  • Knowledge-base-assisted SE: Existing knowledge-base-assisted extraction methods rely on large labeled datasets and well-trained networks, limiting them to systems with unvarying semantic information.They are therefore described as ineffective when semantics or communication context vary over time.
  • Emergent communication: Emergent communication learns semantic and goal-oriented representations through iterative interaction among agents, while contextual-reasoning models report shorter semantic representations with high reliability.The survey notes that many emergent-communication studies still focus on simple, task-specific settings.

E. Some specific SE

The paper examines semantic extraction methods for semantic-aware communication and contrasts their mechanisms, benefits, and limitations across cooperative multi-agent tasks.

  • Semantic-aware communication: In federated deep reinforcement learning, semantic relatedness builds a knowledge graph that selects similar source agents for target-agent training.Semantic relatedness is defined from the source agent’s average return in the target environment over limited training episodes.
  • RL-based SE: Reinforcement-learning semantic extraction converts decoding into a recurrent procedure so specialized semantic metrics can guide training directly.Its benefits include integrating metrics such as BLEU and AoI, while environmental interaction increases training complexity and limits applicability mainly to sequence-generation tasks.
  • KB-assisted SE: KB-assisted semantic extraction stores task-related semantic units and their importance, transmitting selected information according to the knowledge base and channel states.It supports flexible task-specific extraction and SemCom-aware resource allocation, but is limited to non-real-time services and requires computation-intensive knowledge-base construction.
  • Semantic-native SE: Emergent communication learns semantic information and background knowledge through interaction and feedback rather than relying on an existing database.This approach converts passive learning into active learning, but training is time-consuming and convergence is difficult to ensure.
  • Semantic-aware communication: For UAV collision avoidance, semantic information is represented as attention weights between observable agents and supplied to the actor model.Using semantic information instead of raw state data significantly improves training efficiency.

F. Lessons learned summary

The lessons learned emphasize that semantic extraction can improve efficiency and task adaptation, but current methods remain constrained by error floors, training demands, interpretability, and scenario-specific designs.

  • DL-based SE: Attention mechanisms enhance semantic extraction by capturing long-range dependence and removing redundant information through layered re-aggregation.Deep-learning methods can process whole raw inputs and re-extract important information at different layers.
  • DL-based SE: Deep-learning semantic communication has an unavoidable error floor, making it often suboptimal under ideal channel conditions.Joint end-to-end training also leaves semantic extraction and recovery as black-box processes with limited explainability and interpretability.
  • RL-based SE: Reinforcement-learning extraction models sentence decoding recurrently, allowing training to learn correlations among words.The approach naturally fits sequence generation, but recurrent image decoding has unclear necessity and increases decoding time.
  • RL-based SE: 3% accuracy improvement in the middle SNR region is reported for RL-based SemCom over its DL-based counterpart using non-differentiable semantic-metric optimization.The improvement comes with higher training complexity from environment interactions, and feasibility for more complicated language models remains open.
  • KB-assisted SE: KB-assisted extraction is suited to non-real-time services and stable data sources because its knowledge base is computation-intensive to construct and difficult to update frequently.The method also depends heavily on deep-learning models and requires synchronization before communication begins.
  • KB-assisted SE: KB-assisted extraction improves multi-task efficiency by selecting task-related semantic information and avoiding repetitive extraction of raw data.Its knowledge base can also support resource allocation by recording each semantic unit’s size and importance against delay and reliability requirements.
  • Semantic-aware communication: Semantic-aware extraction derives cooperation-relevant information from task properties and agent behavior rather than directly from raw source data.This makes extraction more tailored and difficult to unify, although it is expected to support task-oriented communication.
  • Communication-related challenges: SemCom and conventional communication share constraints from unpredictable channels and limited transmission and processing resources.AI can jointly design source and channel coding, but current AI-based methods lack explicit mathematical explanations.

1) Varying fading channel:

SemCom training must model changing fading channels and uncertain SNR while preserving semantic information under wireless errors. Fixed and generative channel models, adaptation mechanisms, and error correction improve robustness, but generalized solutions remain open.

  • Varying fading channel: Fixed channel-layer schemes use a predetermined fading model throughout SemCom training, including erasure, AWGN, Rayleigh, and Rician channels.Erasure channels map binarized encoder outputs from {-1,1} to {-1,0,1} according to a preset drop probability.
  • Varying fading channel: Generative channel-layer schemes use GAN components to model dynamic channel behavior during training.The generator produces samples resembling real channel data, while the discriminator distinguishes real from generated samples.
  • Varying fading channel: Two-phase training first uses a suitable channel model and then fine-tunes the receiver over the actual channel, achieving lower BLER than without fine-tuning.The optimal training solution for different wireless environments remains unresolved.
  • SNR uncertainty: Fixed-SNR training can produce higher MSE loss at lower tested SNR, leaving broad applicability across SNR ranges uncertain.An SNR-adaptive mechanism injects estimated receiver SNR into channel-feature processing, while channel-wise soft attention scales features by SNR.
  • Bit errors: HARQ combined with joint source-channel coding lowers word and sentence error rates when BER is larger than 0.06.The scheme requests retransmission for uncorrectable code blocks and uses semantic similarity detection between original and estimated sentences.
  • Bandwidth resource: Semantic-relatedness-aware agent selection improved system performance by 83% over a baseline that ignored semantic relatedness.A knowledge graph captures structural and semantic relatedness for bandwidth-limited collaborative training, although dynamic allocation for semantic-content transmission remains insufficiently studied.

2) Energy resource:

SemCom resource and device design must account for semantic importance, energy constraints, heterogeneous hardware, and changing wireless links. Existing approaches include semantic valuation, model compression, multi-agent allocation, and receiver adaptation, but several dynamic settings remain underexplored.

  • Energy resource: Energy allocation can prioritize data containing richer semantic information, while semantic metrics can assess the quality of harvested energy.A semantic valuation function has been used for energy-harvesting IoT devices transmitting text to a hybrid access point.
  • Energy resource: Semantic metrics for energy-resource allocation remain at an early stage, with UAV-aided simultaneous wireless information and power transfer networks still unstudied.The cited boundary concerns SemCom networks requiring energy-harvesting devices.
  • Device capacity: Joint pruning and quantization compresses semantic models by zeroing less significant weights and retaining weights above a pruning threshold.The approach targets devices with limited computing capability.
  • Connections among IoT devices: A multi-agent deep Q-network optimizes semantic video understanding accuracy through spectrum reuse among V2I and V2V links.The reward combines V2I object-detection accuracy and V2V average transmission rate.
  • Connections among IoT devices: Meta-training lets a receiver adapt decoder parameters to unknown channel conditions and achieves lower BLER than conventional training when more than one pilot frame is sent during testing.The approach learns an adaptation rule during meta-training and self-optimizes during testing.
  • Changing resources: Variable-length semantic encoding is needed to address fluctuating spectrum and transmit-power resources in multi-user SemCom networks.The paper compares this need with scalable video coding and multiple description coding in conventional communications.

2) Lessons learned for limited network resources:

SemCom evaluates meaning, timeliness, and task effectiveness rather than only bit-level transmission. Its metrics span semantic error, freshness, and value of information, but remain task-specific and often difficult to integrate into learning systems.

  • Limited network resources: SemCom resource allocation targets accurate task-related semantic information, requiring task-requirement analysis and joint optimization beyond bit transmission.Semantic information introduces a new perspective for resource-allocation design in 6G networks.
  • Limited network resources: Device heterogeneity in capacity and communication environment prevents directly deploying models trained for high-performance devices onto smaller devices.The paper identifies device capacity and wireless environment as the two main heterogeneity dimensions.
  • Semantic metrics: SemCom performance evaluation must address enhanced interlayer coupling through semantic error, AoI, and VoI metrics.These are presented as the three basic metric types and their combined forms.
  • Semantic metrics: Error-based semantic metrics compare intended and understood meaning rather than treating all bits and symbols as equally important.Available semantic metrics are task-specific, and a general metric has not yet been established.
  • Error-based metrics: BLEU compares candidate and reference n-gram matches, while CIDEr measures consensus similarity against human-written descriptions.Sentence similarity uses BERT-based representations to compare transmitted and recovered sentence meaning.
  • Error-based metrics: Sentence similarity is closer to the SemCom paradigm than BLEU and CIDEr because BERT is sensitive to polysemy and whole-sentence meaning.BLEU and CIDEr remain primarily word-difference measures.
  • Metric limitations: Non-differentiable semantic metrics complicate deep-learning semantic extraction, while BERT embeddings increase training resource consumption and hinder generalization.Existing deep-learning pipelines therefore continue to use cross-entropy loss.

2) Semantic metrics for audio data:

Audio and visual SemCom still rely largely on signal or perceptual similarity rather than semantic understanding, while freshness and information value add time-sensitive dimensions. Combined metrics such as AoII address multiple attributes, but metric design remains incomplete.

  • Semantic metrics for audio data: Audio semantic similarity is commonly interpreted as intelligibility, the ease with which receivers understand decoded audio.Existing audio SemCom studies evaluate this dimension using signal-processing metrics.
  • Semantic metrics for audio data: SDR measures speech recovery using L2 error, whereas PESQ models perceptual speech quality across network conditions.PESQ incorporates short-term human perceptual memory but still evaluates transmission accuracy rather than semantic meaning.
  • Semantic metrics for audio data: Audio SemCom lacks metrics that evaluate semantic understanding; existing deep-learning semantic extraction mainly uses MSE.Text-style semantic measures such as BERT and BLEU remain to be studied for audio.
  • Semantic metrics for visual data: Visual SemCom lacks general human-perception-aligned metrics and commonly uses shallow measures such as PSNR and SSIM.Visual similarity is context-dependent and may require high-order structure captured by deep-learning features.
  • AoI-based semantic metrics: AoI quantifies information staleness using the elapsed time since a packet’s timestamp, adding freshness to semantic performance evaluation.Time sensitivity matters for location tracking, control, and situational awareness.
  • AoI-based semantic metrics: AoI disregards recovered-data validity, so it can prioritize useless updates and waste resources when only abnormal source states matter.Time-average age and peak age can be selected according to the stochastic environment.
  • VoI-based semantic metrics: VoI measures information’s importance for decisions, such as abnormal temperature states in control or extracted features in image classification.In many cases, VoI is known only after the relevant task outcome is observed.
  • Combined semantic metrics: AoII combines content accuracy and timeliness by measuring the impact of prolonged inaccurate states on semantic recovery.It accounts for both transient and duration-dependent effects on the communication goal.

E. Lessons learned summary

The survey reviews semantic metrics, SemCom applications, and implementation challenges across several 6G scenarios. It emphasizes that semantic evaluation remains immature and that practical systems face estimation, resource, and knowledge constraints.

  • Lessons learned: Semantic performance evaluation remains less developed than semantic extraction and transmission research.Semantic metrics move beyond bit- and symbol-level comparisons by considering conveyed meaning and information relevance.
  • Lessons learned: Existing semantic metrics are intrusive, requiring reference signals that are often unavailable in real-world communications.Their complex, non-differentiable forms also restrict their use during model training.
  • Applications: SemCom applications span intelligent transport, federated learning, and UAV networks, where systems extract or transmit task-relevant information under resource constraints.The reviewed examples address vehicle data, model updates, and UAV communication links.
  • Applications: UAV swarm navigation achieved 6.5x lower latency at a target 10^-7 error rate than a state-of-the-art CTDE-based method.

4) Extended reality:

SemCom is presented as a way to support immersive 6G services by extracting and transmitting information relevant to applications rather than raw content. The proposed IE-SC architecture organizes this process across semantic layers coordinated by a semantic intelligence plane.

  • Extended reality: SemCom can support XR-based Metaverse access by filtering device-tracked movements, gestures, and speech before transmission.This is intended to save bandwidth and reduce computing latency at the XR server.
  • Extended reality: Traditional content-blind communication wastes bandwidth and computing resources for communication-and-computation-intensive services.SemCom instead uses understanding before transmission and adapts semantic extraction to network conditions.
  • Collaborative robots: In collaborative-robot communication, L-DeepSC compresses required interaction data to 2.5% of the amount needed by a traditional method under low SNR.
  • SemCom-empowered 6G architecture: The IE-SC architecture replaces the seven-layer OSI organization with S-AI, S-NP, and S-PB layers coordinated through a semantic intelligence plane and S-IF.
  • SemCom-empowered 6G architecture: The semantic intelligence plane extracts semantic information, manages shared background knowledge, and evaluates performance for network-layer decisions.

2) Semantic application-intent layer:

The semantic application-intent layer translates user or application purposes into representations and policies that guide downstream semantic communication. Its design connects intent analysis with semantic interaction and protocol adaptation.

  • Semantic application-intent layer: The S-AI layer analyzes communication intent and translates it into network deployment, configuration, or control policies.
  • Semantic application-intent layer: Intent mining extracts, analyzes, aggregates, and synthesizes original intents, which can produce different communication processes for the same image.
  • Semantic application-intent layer: Intent decomposition converts mined intent into sub-intents representing attention to text elements or importance of image features.
  • Semantic application-intent layer: Semantic representation encodes the sub-intent set and passes it to the semantic intelligence plane to facilitate semantic extraction.
  • Semantic network-protocol layer: The S-NP layer serves upper-layer intents through semantic interaction, knowledge sharing, protocol analysis, protocol formation, and SI encapsulation.
  • Semantic physical-bearing layer: Semantic physical-bearing modules convert semantic information into physical signals using separate semantic coding, channel-aware coding, or joint source-channel coding.

A. Interpretability and explainability of SE

The survey identifies interpretability, resource demands, semantic caching, implicit meaning, channel complexity, and security trade-offs as open issues for practical SemCom. These concerns constrain predictability, deployment, and evaluation beyond transmission accuracy.

  • Interpretability and explainability: Black-box semantic extraction models behave unpredictably under uncertain inputs, limiting social acceptance, practical use, and principled optimization.
  • Interpretability and explainability: Interpretability can reveal why semantic models make decisions, exposing strengths and weaknesses while guiding improvements.
  • System resources: Training and updating semantic extraction models require substantial storage, computation, and communication resources, especially as contexts evolve.
  • Semantic caching: Semantic caching must prioritize whether cached information can be inferred accurately by the requester, not merely whether data content has a high hit rate.
  • Implicit SemCom: Most prior SemCom transfers explicit information, while implicit meaning requires receivers to infer context-dependent interpretations that lack directly specified utility functions.
  • Implicit SemCom: GAML improved accuracy by 20% over genetic-algorithm reasoning, while additive and linear inference recovered missing information with 76% and 48% accuracy, respectively.
Loading 2207.00427v2…