Source-linked AI summary
Resource allocation for text semantic communications
Lei Yan, Zhijin Qin, Rui Zhang, Yongzhao Li, Geoffrey Ye Li
TL;DR
Resource allocation for semantic communications remains unexplored, while bit-based spectral efficiency does not apply because bits reflect source-symbol statistics rather than semantic information. The paper defines semantic spectral efficiency for text communication, optimizes channel assignment and transmitted semantic symbols, develops a conversion method for fair comparison, and reports higher S-SE for semantic communication than conventional systems.
Problem
Resource allocation for semantic communications requires a semantic measure of information content and spectral efficiency because conventional bit-based measures do not apply.
Method
The paper defines S-SE, formulates resource allocation over channel assignment and transmitted semantic symbols, and converts bit-based SE to S-SE for comparison.
Results
Simulation results verify the proposed resource allocation model and report higher S-SE for semantic communication than conventional communication systems, including 4G and 5G.
Takeaways & Limitations
For text transmission, semantic-aware resource allocation provides a basis for measuring and optimizing communication efficiency in the semantic domain.
Abstract
from arXiv · showhide
Semantic communications have shown its great potential to improve the transmission reliability, especially in the low signal-to-noise regime. However, resource allocation for semantic communications still remains unexplored, which is a critical issue in guaranteeing the semantic transmission reliability and the communication efficiency. To fill this gap, we investigate the spectral efficiency in the semantic domain and rethink the semantic-aware resource allocation issue. Specifically, taking text semantic communication as an example, the semantic spectral efficiency (S-SE) is defined for the first time, and is used to optimize resource allocation in terms of channel assignment and the number of transmitted semantic symbols. Additionally, for fair comparison of semantic and conventional communication systems, a transform method is developed to convert the conventional bit-based spectral efficiency to the S-SE. Simulation results demonstrate the validity and feasibility of the proposed resource allocation method, as well as the superiority of semantic communications in terms of the S-SE.
I. INTRODUCTION
The paper addresses the lack of semantic-domain spectral-efficiency measures and resource-allocation methods for semantic communications. It defines S-SE, optimizes channel and semantic-symbol allocation, develops a conversion for fair comparison, and evaluates the approach through simulations.
- Conventional bit-based spectral efficiency does not represent semantic information, so resource allocation must be reconsidered from the semantic perspective.
- Prior semantic-information studies established theoretical foundations but did not quantify spectral efficiency in the semantic domain or provide practical implementation guidance.
- DeepSC enables calculable semantic-domain efficiency analysis by extracting and delivering text meaning with deep-learning-based semantic communication.
- The paper defines semantic spectral efficiency and formulates resource allocation to maximize overall S-SE through channel assignment and transmitted semantic-symbol selection.
- A transformation method converts conventional bit-based spectral efficiency into S-SE for fair comparison with semantic communication systems.
- Simulations verify the proposed resource-allocation model and report semantic communication superiority in S-SE.
II. SYSTEM MODEL
The system models a cellular network in which users transmit text through DeepSC over assigned fading channels. Each user's sentence is mapped to a variable-length semantic-symbol vector, while channel assignment is constrained to one channel per user and one user per channel.
- The network consists of a base station and multiple users, with DeepSC adopted for semantic transmission.
- Each user maps a sentence of length L_n into a semantic-symbol vector whose length is k_nL_n and varies with sentence length.
- The parameter k_n denotes the average number of semantic symbols transmitted per word for user n.
- Available channels share bandwidth W, and binary assignment variables allocate each channel to at most one user while limiting each user to at most one channel.
- The channel model includes large-scale fading, small-scale Rayleigh fading, user transmit power, channel gain, and noise power spectral density.
C. DeepSC Receiver
The receiver decodes the received signal and evaluates reconstructed text using BERT-based semantic similarity. This metric supports the paper's semantic-efficiency formulation and resource-allocation objective.
- The base station applies channel decoding followed by semantic decoding to estimate each user's transmitted sentence.
- Text-transmission performance is evaluated using semantic similarity between the original and reconstructed sentences.
- The semantic similarity metric uses a pretrained Sentence-BERT model and measures the distance between sentence-level semantic representations.
- The paper defines S-SE and formulates semantic-aware resource allocation as an optimization problem over channel assignment and transmitted semantic-symbol count.
A. Semantic Spectral Efficiency
The paper introduces semantic-level transmission metrics because bit-based spectral efficiency does not represent semantic information. For text transmission, S-SE measures successfully transmitted semantic information per unit bandwidth and incorporates semantic similarity, source statistics, and symbol allocation.
- Bit-based spectral efficiency cannot measure semantic information because transmitted bits reflect source statistics rather than source meaning.
- Semantic transmission rate measures effectively transmitted semantic information per second in suts/s.
- Semantic spectral efficiency measures successfully transmitted semantic information per unit bandwidth in suts/s/Hz.
- The framework uses expected semantic information and sentence length for long-term text transmission rather than individual-sentence values.
- S-SE incorporates semantic similarity, which depends on the semantic-symbol allocation kn and channel conditions γn,m.
B. Problem Formulation
The resource allocation model maximizes overall S-SE across users by jointly selecting channel assignments and the average number of transmitted semantic symbols per word. The formulation includes channel, symbol-count, semantic-similarity, and user-efficiency constraints.
- The model maximizes the overall S-SE of all users.
- Channel assignment and the average semantic-symbol count per word kn are optimized jointly.The symbol-count variable is intended to increase information carried per symbol while preserving transmission reliability.
- The optimization problem includes channel-assignment constraints and bounds on the average number of semantic symbols per word.
- The formulation also enforces minimum semantic similarity ξth and minimum user S-SE Φth.
C. The Optimal Solution
The solution removes the source-dependent constant from the optimization, obtains semantic-similarity mappings through DeepSC experiments over AWGN, and decomposes the problem into searchable channel and assignment subproblems.
- The ratio I/L is constant for a given source type, so it can be omitted without affecting resource optimization.
- The mapping between semantic similarity ξn,m and (kn, γn,m) is obtained by running DeepSC over an AWGN channel.
- Orthogonal cellular links allow the optimization to be decoupled into two equivalent independent problems.
- Exhaustive search obtains candidate-channel values using the lookup-table semantic-similarity constraints, while the Hungarian algorithm solves the resulting bipartite maximum-matching problem.
IV. SIMULATION RESULTS AND COMPARISON
The simulations evaluate both the proposed semantic-aware allocation model and the relative S-SE of semantic versus conventional communication systems. A transform method enables the latter comparison by converting conventional bit-based SE into S-SE.
- The simulations compare the proposed resource allocation model against the conventional allocation model.
- The study compares semantic and conventional systems using S-SE to assess semantic communications.
- A transform method converts conventional bit-domain spectral efficiency to S-SE while accounting for source coding.
A. The Transform Method for Fair Comparisons
The paper transforms conventional bit-based spectral efficiency into semantic spectral efficiency by accounting for source coding and bit transmission, enabling fair comparisons with semantic systems.
- Conventional bits are treated as approximate semantic symbols, although they may carry less semantic information than DeepSC symbols.
- The transform uses transmission rate C_n,m, bandwidth W, source-coding factor µ, and error-related factor ξ_n,m to derive equivalent S-SE.µ represents the average number of bits per word and reflects source-coding compression; R_n,m = C_n,m/W is conventional SE.
- The source coding and bit transmission processes are both included, allowing fair comparisons between conventional and semantic communication systems.
B. Benchmarks
The benchmarks comprise an ideal Shannon-limit system and practical 4G and 5G systems, whose bit-based efficiencies are transformed into S-SE for comparison.
- The proposed semantic resource allocation is compared with one ideal benchmark and two widely deployed practical benchmarks: 4G and 5G.
- The ideal system assumes error-free transmission and achieves R_n,m = log2(1 + γ_n,m).
- The 4G benchmark obtains channel quality indicators from measured SNR and maps them to achievable SE using 3GPP TS 36.213.
- The 5G benchmark similarly uses measured-SNR CQI and 3GPP TS 38.214 to determine achievable SE, then applies the transform method for S-SE optimization.
- The benchmark optimization uses the same method introduced for the proposed S-SE resource allocation problem.
C. Simulation Results
The proposed semantic-aware resource allocation achieves higher S-SE than the conventional model, while semantic communication generally outperforms conventional systems under the studied conditions. Its advantage depends on transmit power, channel availability, and the conventional source-coding transform factor.
- Resource allocation comparison: The conventional resource allocation model produces lower S-SE than the proposed model for all examined k_n values, reaching zero at k_n = 3.At k_n = 3, semantic similarity falls below the threshold.
- Resource allocation comparison: S-SE rises rapidly as channels increase from 1 to 5, then grows more slowly from 5 to 10 as users select channels with higher SNR.More channels serve more users initially; later gains are smaller because channel selection improves.
- System comparison: The semantic communication system outperforms all conventional communication systems in the comparison shown in Fig. 4.
- System comparison: With increasing transmit power, practical systems’ S-SE first increases and then approaches an upper bound, while the ideal system continues increasing rapidly.The semantic system has a larger upper bound than 4G and 5G because of stronger data compression.
- Transforming factor: The semantic system is stable across the transforming factor, exceeds 4G and 5G above 19 bits/word, but underperforms the ideal system below approximately 27 bits/word.For conventional systems, S-SE decreases as the transforming factor increases because maximum SE is fixed and S-SE divides SE by µ.
V. CONCLUSION
The paper defines semantic-domain efficiency measures and formulates semantic-aware resource allocation for text transmission using DeepSC. Simulations show that semantic communication can exceed 4G and 5G, with the outcome depending on conventional source coding, while broader multi-task allocation remains future work.
- V. CONCLUSION: S-R and S-SE are defined using the DeepSC model to measure semantic communication efficiency, and resource allocation is optimized over all users.The allocation problem covers channel assignment and the number of transmitted semantic symbols.
- V. CONCLUSION: For text transmission, semantic communication achieves higher S-SE than 4G and 5G when conventional source coding maps a word to more than 19 bits on average.
- V. CONCLUSION: With 10 dBm transmit power and more than 27 bits required per encoded word, semantic communication outperforms the ideal system.
- V. CONCLUSION: Resource allocation for multiple intelligence tasks, including single-modal and multimodal tasks, remains an open direction for future investigation.