Source-linked AI summary
Semantic-Aware Sub-Band Allocation for Terahertz Communications
Fatima Ismail, Hadi Sarieddeen, Jihad Fahs
TL;DR
The paper addresses semantic-aware allocation over frequency-selective THz bands, where sentence-dependent nonlinear SNR responses make assignment non-separable. It uses an SBERT surrogate and imitation-trained neural scheduler to approximate oracle allocation, achieving consistent benchmark gains and near-oracle performance under realistic channels.
Problem
Frequency-selective THz allocation methods ignore semantics, while semantic-aware allocation methods overlook pronounced THz channel selectivity, leaving their joint treatment unexplored.
Method
An SBERT surrogate predicts sentence-level semantic fidelity from embeddings and SNR, while a neural scheduler learns oracle assignments generated from surrogate utility through imitation learning.
Results
The proposed scheduler consistently outperforms practical benchmarks, approaches oracle performance, and maintains a utility gap below 0.25% across tested K values.
Takeaways & Limitations
Content-dependent allocation that aligns semantic importance with heterogeneous THz sub-band quality improves reliability beyond traditional bit-centric resource allocation.
Abstract
from arXiv · showhide
This paper studies semantic-aware sub-band al- location for terahertz (THz) communication systems, where frequency-selective molecular absorption creates highly non- uniform sub-band qualities. Unlike conventional formulations, semantic fidelity depends nonlinearly on the signal-to-noise ratio (SNR) and is also sentence-specific, leading to a non-separable assignment problem that is generally not solvable using simple ordering-based policies. To address this, we use a sentence-BERT (SBERT)-based surrogate model that predicts semantic fidelity from the sentence embedding and sub-band SNR. We propose an importance-aware scheduler that assigns sentences to sub- bands based on their semantic contribution using an oracle utility function that captures importance-weighted semantic similarity across sentence-sub-band pairs. A neural scheduler is then trained through imitation learning to approximate the oracle policy at more than 200x lower runtime than full DeepSC- based oracle evaluation. Integrated with a deep-learning-enabled semantic communication (DeepSC) system, the proposed method consistently outperforms all benchmark schemes and approaches oracle-level performance under realistic THz channel conditions.
I. INTRODUCTION
The paper addresses semantic-aware sub-band allocation for THz systems, where molecular absorption creates heterogeneous channel quality and sentence-specific semantic responses make conventional allocation inadequate. It proposes surrogate-assisted, importance-aware scheduling integrated with DeepSC.
- Motivation: THz molecular absorption creates frequency-selective sub-bands with highly heterogeneous channel qualities.Existing THz allocation methods are bit-centric, while semantic-aware methods typically overlook pronounced THz frequency selectivity.
- Approach: The paper formulates semantic-aware THz allocation by jointly modeling sentence importance and frequency-selective channel conditions.The formulation is built for a DeepSC-based semantic communication system operating over distinct THz sub-bands.
- Research gap: Semantic fidelity is nonlinear in SNR and sentence-dependent, producing a non-separable assignment problem beyond simple sorting policies.The utility of a sentence–sub-band pairing depends jointly on sentence content and experienced SNR.
- Approach: An SBERT-based surrogate approximates semantic fidelity from sentence embeddings and SNR, avoiding repeated full DeepSC evaluations.The surrogate supports efficient utility estimation for sentence–sub-band assignment.
- Contribution: A neural scheduler trained by imitation learning approximates oracle assignments at low complexity and consistently improves over benchmark schemes under realistic THz channels.The oracle assignments are generated from the surrogate-assisted optimization process.
III. PROBLEM FORMULATION
The system serves multiple users whose sentences are transmitted through orthogonal THz sub-bands with distinct SNRs. Semantic fidelity is measured by SBERT cosine similarity and depends jointly on the sentence and assigned SNR.
- System model: A transmitter serves K users, with each user providing one sentence per time slot.The system assumes one-to-one assignment between users and sub-bands.
- Channel model: The THz window contains S orthogonal sub-bands whose SNRs differ because of path loss and frequency-selective molecular absorption.Inter-user interference is neglected, and S = K is used for one-to-one assignment.
- Transmission model: Each sentence is encoded with DeepSC, transmitted over its assigned sub-band, and reconstructed at the receiver.The reconstructed sentence is used to evaluate semantic fidelity.
- Semantic metric: Semantic fidelity is measured using SBERT cosine similarity and represented as m = g(γs, v), a nonlinear function of SNR and sentence content.Because the reconstruction depends on both γs and v, semantic quality is not determined by channel quality alone.
A. Semantic Importance and Assignment Model
The assignment model weights sentence-level semantic fidelity by importance and enforces one-to-one sentence–sub-band matching. Since semantic similarity varies nonlinearly with SNR across sentences, the resulting utility depends on the assignment structure.
- Semantic importance: Each sentence receives an importance score πi ∈ [0, 1] representing its semantic relevance.The score is computed from token-level MARS importance weights using average and maximum token importance.
- Semantic importance: The sentence-level importance combines global average and peak token importance, with α balancing the two contributions.The paper sets α = 0.5, while other values may apply depending on the task.
- Assignment model: The binary assignment matrix X assigns every sentence to exactly one sub-band and every sub-band to exactly one sentence.This one-to-one constraint defines the feasible allocation for each time slot.
- Assignment model: The semantic utility captures importance-weighted fidelity under X, and the allocation objective selects the assignment maximizing that utility.The formulation evaluates semantic performance jointly with sentence importance and sub-band conditions.
- Problem structure: For each sentence, semantic similarity is non-decreasing but nonlinear in SNR and dependent on the sentence.This sentence-specific response is the key structural property of the optimization problem.
1) Pairwise Exchange Analysis:
Pairwise exchange analysis shows why sentence–sub-band assignment cannot generally be reduced to sorting importance scores against SNRs. Different sentences can obtain different marginal benefits from the same SNR improvement.
- Pairwise Exchange Analysis:: Different sentences benefit differently from SNR improvements, making the marginal gain of a higher-quality sub-band sentence-dependent.Optimal allocation is therefore driven by importance-weighted marginal gains rather than simple ordering.
- Pairwise Exchange Analysis:: Under a common semantic response g(γ, vi) = h(γ) for all sentences, greedy sorting becomes globally optimal.The sufficient condition requires h(γ) to be non-decreasing.
- Pairwise Exchange Analysis:: When importance scores and sub-band SNRs are oppositely ordered, exchanging the two sub-bands does not decrease utility.Repeated exchanges yield the sorted assignment as the global optimum under the shared-response condition.
IV. SYSTEM ARCHITECTURE
The system combines SBERT-based semantic-fidelity estimation with neural sub-band scheduling for DeepSC over frequency-selective THz channels. Surrogate errors perturb utility estimates, while oracle-loss analysis links prediction accuracy to assignment reliability.
- Pipeline: The end-to-end pipeline uses offline surrogate and scheduler training, online assignment inference, and DeepSC-based semantic transmission.The surrogate and Hungarian solver are used during training, while the scheduler performs assignment inference online.
- Surrogate: The SBERT surrogate estimates semantic similarity from sentence embeddings and SNR, avoiding repeated full DeepSC evaluations.It approximates g(γ, v) for efficient utility evaluation.
- Scheduler: The neural scheduler maps sentence semantics, importance, and channel state to one-to-one sub-band assignments.Its permutation-valued output supports low-complexity allocation inference.
- Surrogate Error: Surrogate prediction errors enter utility entrywise as π_i e_i,s, directly controlling aggregate utility perturbation.The error is defined as e_i,s = ˆg_i(γ_s) − g_i(γ_s).
- Surrogate Error: The expected surrogate-induced oracle loss is bounded using the worst-case expected pairwise error, linking surrogate accuracy to oracle-label reliability.The analysis defines η_i,s as expected absolute error and η_max as their maximum.
V. STUDIED ALGORITHM
The studied algorithm trains an SBERT surrogate first, then uses its utility estimates and Hungarian assignments to train a scheduler by imitation learning.
- Training Stages: The offline algorithm has separate surrogate-training and scheduler-training stages.These stages correspond to Algorithms 1 and 2.
- Surrogate Training: Surrogate training forms embedding–SNR inputs and predicts semantic similarity with a learned model.Algorithm 1 iterates over minibatches and optimization epochs.
- Oracle Construction: For each frame, the algorithm constructs the surrogate utility matrix ˆU_i,s = π_i ˆm_i,s and obtains oracle assignments with the Hungarian algorithm.The utility matrix uses sentence embeddings, importance scores, and sub-band SNRs.
- Scheduler Training: The scheduler is trained with cross-entropy to imitate Hungarian-based oracle assignments.The trainable parameters are updated using gradient descent on the assignment targets.
- Inference: At inference, the surrogate and Hungarian solver are discarded, and the scheduler produces assignments in a single forward pass.This separates expensive oracle construction from deployment-time scheduling.
VI. RESULTS AND DISCUSSION
The evaluation compares importance-aware allocation with random, importance-only, SNR-only, and utility-proxy baselines under realistic indoor THz conditions. It also examines surrogate behavior through figures covering semantic fidelity and oracle gaps.
- Baselines: The evaluation compares random assignment, PiOnly, SNROnly, greedy, proxy Hungarian, and oracle Hungarian schemes.The baselines isolate the effects of semantic importance, SNR, and their combinations.
- Simulation Setting: The indoor THz scenario spans molecular-absorption peaks from 380–450 GHz over a 5 m link with S = K frequency-selective sub-bands.The channel includes frequency-dependent attenuation and re-applies scheduling each frame.
- Surrogate Accuracy: Figure 2 plots average true and predicted SBERT semantic fidelity against sub-band SNR.The figure compares the surrogate prediction with the true semantic-fidelity curve.
- Oracle Gap: Figure 3 estimates the normalized surrogate-induced oracle gap E[U(X⋆)−U(ˆX⋆)].The quantity measures the utility difference between true-utility and surrogate-utility oracle assignments.
- Simulation Setting: The simulations use clustered multipath propagation and Europarl sentences with 80k/10k training and test samples.The channel model includes Poisson-distributed clusters and Gaussian angular spreads.
A. Surrogate Accuracy
The SBERT surrogate closely follows true semantic similarity across SNR, with small average prediction error but larger worst-case deviations at very low SNR. The resulting oracle gap decreases as transmit power increases.
- Prediction Accuracy: The surrogate closely tracks true mean SBERT similarity across SNR and achieves near-zero mean error on 56,250 test samples.The reported mean error is ˆμ_e = −2.45×10^-4.
- Prediction Accuracy: 5.33 × 10^-2 mean absolute error indicates small typical surrogate prediction error.Because utility perturbations scale with π_i e_i,s and 0 ≤ π_i ≤ 1, typical utility distortion is limited.
- Prediction Accuracy: η_max ≈ 0.20 is larger because of outlier sentence–sub-band pairs at very low SNR, making the theoretical bound conservative.The average error remains much smaller than this empirical worst-case value.
- Oracle Gap: The normalized oracle gap is largest at low transmit power and decreases rapidly with P_t.The surrogate-based oracle approaches the true oracle in the reliable semantic operating regime.
B. Importance-Weighted Semantic Utility
For K = 15, the proposed scheduler outperforms practical baselines across transmit powers and closely approaches the oracle, with especially large gains at low power. It prioritizes important sentences while retaining acceptable performance for less important content.
- The comparison evaluates allocation policies at K = 15, while K = {5, 10} exhibit the same trends and are omitted for clarity.
- At low Pt, the proposed scheduler delivers the largest gains over practical baselines and closely approaches the oracle.Frequency selectivity and nonlinear semantic degradation make allocation especially important in this regime.
- As Pt increases, all methods converge because semantic fidelity saturates at high SNR, while the scheduler retains a consistent advantage.
- Importance prioritization: For top-10% sentences, the scheduler achieves near-oracle semantic fidelity across all Pt, with greatest gains at low power.This reflects prioritization of critical content using both semantic importance and channel quality.
- Fairness: For bottom-10% sentences, performance remains acceptable, indicating a balance between prioritization and fairness.At low Pt, lower performance follows from allocating better resources to important sentences.
C. Scalability Analysis
The learned scheduler scales across tested problem sizes while maintaining near-oracle utility and stable assignment latency. It is substantially cheaper than oracle construction, though simple heuristics remain faster.
- The scheduler maintains a utility gap below 0.25% relative to the surrogate-assisted Hungarian oracle for K = S ∈{5, 10, 15}.The learned policy generalizes across the tested problem sizes.
- Assignment latency remains near 0.49 ms across all tested K, while surrogate-assisted Hungarian latency is 2–3× higher.These values refer to online assignment only.
- Latency: Oracle construction is over 200× slower because it requires explicit utility-matrix construction and combinatorial assignment.
- Latency: The scheduler uses a single feed-forward pass, offering near-oracle performance at orders-of-magnitude lower complexity than the oracle.Its runtime is higher than that of simple heuristics.
- The proposed system integrates DeepSC with a lightweight neural policy trained through surrogate-assisted imitation learning.An SBERT-based surrogate enables efficient fidelity estimation for the training pipeline.