Phones and vehicles increasingly send what users write or say to an edge server, and that server runs models that read the content. One such task is multi-label emotion classification: deciding which of several emotions, such as anger, disgust, optimism or sadness, are present together in a sentence. Recent "semantic communication" systems skip exact bit recovery. A small encoder on the device maps the content directly to radio symbols, and a model at the receiver decodes the task answer from those symbols. These systems are tested almost entirely under the quasi-static (block-fading) assumption, which treats the radio channel as frozen for the length of a packet. At vehicular speeds that assumption fails. The channel rotates and drifts within a single frame, so symbols sent just after a reference signal (a pilot) are received cleanly, and later symbols are received through an increasingly outdated channel estimate.
This note proposes one question and one method. The question is whether multi-label emotion inference needs the pilot density that channel-estimation accuracy would demand. Our hypothesis is that it needs fewer pilots, because a label decoder that knows how stale each received symbol is can down-weight unreliable symbols and fall back on known label co-occurrences. We would build a device-to-edge emotion system trained under explicit within-frame channel aging. Its receiver is told each symbol's "age" since the last pilot. We would measure how the pilot spacing that is optimal for the task differs from the spacing that is optimal for channel error. We would also report two side questions: how badly quasi-static-trained systems fail under Doppler, and which emotion confusions the failure creates.
Core Claim
The claim has two parts. First, a semantic emotion system trained on block fading will lose more multi-label accuracy under realistic Doppler than its block-fading evaluation suggests. The loss will concentrate on correlated negative-emotion pairs (anger versus disgust) that are already the weak spot of text-only models, because those decisions depend on small differences in the received features. Second, an encoder trained under a differentiable aging channel, paired with a decoder that receives a per-symbol staleness estimate, will match the block-fading system's F1 at noticeably sparser pilot spacing. The new element is the claim that pilot overhead should be budgeted against label-set accuracy rather than against channel mean-squared error, together with a decoder built to exploit that difference.
Relation to Prior Work
The closest existing system is BR-GG-DeepSC (Multimodal semantic communication system based on graph neural networks). It transmits image and text features end to end for emotion analysis and uses a receiver-side LSTM to correct channel distortion. It differs from this proposal in three ways. It is evaluated only on AWGN and block-faded Rayleigh/Rician channels. It predicts coarse single-label sentiment on MVSA rather than multi-label sets. Its text branch actually used more channel symbols than a conventional pipeline. That paper therefore motivates the proposal by showing that text-side semantic transmission remains unsolved, but it does not test what we test.
SGD-JSCC (Semantics-Guided Diffusion for Deep Joint Source-Channel Coding in Wireless Image Transmission) does separate slow and fast fading. Its pilot-free estimator, however, is restricted to slow fading. Its fast-fading treatment equalizes noise variance across blocks rather than tracking drift within a frame. It also targets image reconstruction and needs about 1.5 s per sample, so it constrains rather than supplies our design. Work on contactless affect sensing (WiFE, micro-Doppler facial emotion) uses channel fluctuations as the sensor itself, which is a different problem.
On the channel side, several papers supply the physics and the trade-off we reuse. The non-stationary MIMO model of A Novel 3D Non-Stationary Multi-Frequency Multi-Link Wideband MIMO Channel Model explains why stationarity breaks at speed. The training-versus-data analysis in Joint Optimization of Power Allocation and Training Duration for Uplink Multiuser MIMO Communications frames the pilot trade-off we re-optimize. That paper works under block fading and with a sum-rate objective; we swap in a task objective and an aging channel. Two papers show that learned receivers can cut pilots or reuse data. Pruning the Pilots reaches −15 dB NMSE with 8 rather than 16 pilots. Detect to Learn tracks 30 km/h channels using decisions as pseudo-pilots. Both still optimize channel or bit accuracy, not a downstream label task. MambaCSP supplies a linear-time channel predictor and the QuaDRiGa UMa setup (10–100 km/h), which we adopt for testing. Its finding that a 1.3B model barely beats a 130M one suggests a small predictor is enough for the staleness estimate.
On the emotion side, the receiver reuses SpanEmo's design (SpanEmo). SpanEmo places emotion-label tokens in the attention context and adds a label-correlation ranking loss. Its ablation found that the label segment mattered more than the loss, which is why we keep label tokens in the decoder. The known failure on anger/disgust motivates our per-pair analysis.
Research Program
The device runs a small text encoder, on the order of a distilled BERT. Its pooled and token features are projected to a fixed number of complex channel symbols, arranged in a frame that has a configurable pilot spacing.
Training passes these symbols through a differentiable time-varying channel. We use a Clarke–Jakes autoregressive fading process with Doppler sampled from 0 up to the 100 km/h range, so the channel changes symbol by symbol. The receiver forms a least-squares channel estimate at each pilot and holds it until the next pilot, which is how CSI aging arises.
A lightweight channel predictor runs at the receiver; we start with a Kalman/AR filter and try a small SSM if needed. It outputs, for each data symbol, an equalized value and a predicted error variance. The edge decoder is SpanEmo-style. Label tokens cross-attend to the received symbol features, and each symbol's predicted error variance is injected as an extra input channel. We call this staleness conditioning. The loss is SpanEmo's BCE plus label-correlation term, optionally plus a channel-use penalty.
The study design has three parts. We sweep pilot spacing, SNR and speed. For each setting we record both the channel NMSE and the multi-label F1, and we locate each metric's optimal pilot spacing. We then compare those optima against the separate-channel baselines below. Testing uses QuaDRiGa 3GPP UMa NLOS traces, which the model never sees in training. This partially addresses the simulation-to-training gap, though not the sim-to-real gap.
Baselines
The naive reference is the quasi-static-trained system. It uses the same encoder, decoder and symbol budget, but is trained on block Rayleigh fading with one pilot block per frame, as in the evaluations of BR-GG-DeepSC. It is then tested on the same aging channels. It shares parameters, data and tuning budget with our system, so any gap is attributable to the training channel. Beating it means higher Micro F1 at every Doppler above the coherence-limited regime, at equal channel uses.
The strongest existing method on the emotion task is SpanEmo itself, run on clean text at the edge. SpanEmo reports Micro F1 0.713 and Jaccard 0.601 on SemEval-2018 E-c (English) (SpanEmo). In our setting this is the noiseless upper bound. The deployable form is a separation pipeline: the device sends the text as 5G-style LDPC-coded QAM with standard adaptive pilot insertion, and SpanEmo runs on whatever bits arrive. We give it the same total channel uses per sentence, including its own pilots, and let it tune pilot spacing and modulation and coding scheme over the same grid. "Beating" it means equal or better F1 at fewer channel uses, or higher F1 at matched uses below the LDPC cliff, while staying within, say, two points of 0.713 at high SNR. For the closest semantic baseline, no reported numeric accuracy for BR-GG-DeepSC on MVSA was found on Lacuna. We would reimplement its text branch and LSTM corrector on our data rather than quote a figure. For GoEmotions, the reference point is the BERT-base macro F1 of 0.46 over 27 classes (GoEmotions).
Three ablations isolate the key ingredient. The first trains on aging channels but removes the staleness input; it tests whether aging-aware training alone accounts for the gain. The second keeps the staleness input but shuffles symbol order at test time; it tests whether a learned feature-to-position ordering matters. The third replaces SpanEmo's label tokens with a flat sigmoid head; it tests whether label-context attention is what absorbs channel noise.
Why not just…?
Why not just run a small classifier on the device and send the 11–28 label bits with heavy coding? This is the most serious threat. For a pure label-output task it may simply win, since a few dozen bits with repetition coding survive almost any channel. Its weakness is classifier quality. A device-sized model gives up accuracy relative to an edge model with label-context attention. The proposal is only worthwhile if the edge model's accuracy advantage exceeds the channel loss of sending features. We include this as a matched-device-compute baseline: the same on-device encoder with a classification head, with labels sent over the same channel. If it matches our system across the speed and SNR sweep, we will report that the task does not justify feature transmission. In that case the task-aware pilot-budget finding would be reframed for settings where the text must also be delivered.
Why not just send the text digitally and classify at the edge? That separation pipeline is the strongest baseline above. The concern from Joint Optimization of Power Allocation and Training Duration is that tracking a fast channel for exact bit recovery costs a large share of energy in training, up to roughly half at low SNR in their block-fading analysis. LDPC also fails abruptly once the post-aging SNR drops below threshold. Whether that cost exceeds the semantic system's loss is exactly what the matched-channel-use sweep measures. We expect separation to win at high SNR and low speed.
Why not just train the end-to-end system on time-varying channels and skip the staleness input? A learned encoder might implicitly put the important features in early, fresh symbols, which would make explicit conditioning redundant. The first ablation tests this directly. If the gap between it and the full system is within seed noise, the contribution reduces to "train on the right channel" plus the pilot-budget analysis, and we would say so.
Proposed Experiments and Decisive Checks
The datasets are SemEval-2018 Task 1 E-c (11 labels, English, Arabic and Spanish) (SemEval-2018 Task 1) and GoEmotions (28 labels). We report Micro and Macro F1, Jaccard, per-pair confusion rates for anger/disgust and other correlated negative pairs, channel NMSE, and channel uses per sentence. Each result is plotted against SNR (−5 to 20 dB), speed (0–100 km/h) and pilot spacing, averaged over at least three seeds.
The first decisive check is the gap between task-optimal and NMSE-optimal pilot spacing. If the F1-optimal spacing is not reliably sparser, the central claim fails.
The second check concerns where the quasi-static system's errors go. We test whether its degradation concentrates on correlated negative pairs more than on uncorrelated labels, which would confirm that channel aging specifically amplifies the label-confusion weakness.
The third check is the three-way comparison at matched channel uses between our system, separation with SpanEmo, and on-device classification with label bits. This settles whether feature transmission is justified at all.
As a secondary audit, we run the Equity Evaluation Corpus from SemEval-2018 through the channel. We would check whether channel noise widens the demographic score gaps that 75–86% of that task's systems already showed.
References
- Multimodal semantic communication system based on graph neural networks
- Semantics-Guided Diffusion for Deep Joint Source-Channel Coding in Wireless Image Transmission
- WiFE: WiFi and Vision based Intelligent Facial-Gesture Emotion Recognition
- Privacy-Preserving Facial Emotion Classification With Visual Micro-Doppler Signatures for Hearing Aid Applications
- A Novel 3D Non-Stationary Multi-Frequency Multi-Link Wideband MIMO Channel Model
- Joint optimization of power allocation and training duration for uplink multiuser MIMO communications
- Pruning the Pilots: Deep Learning-Based Pilot Design and Channel Estimation for MIMO-OFDM Systems
- Detect to Learn: Structure Learning With Attention and Decision Feedback for MIMO-OFDM Receive Processing
- MambaCSP: Hybrid-Attention State Space Models for Hardware-Efficient Channel State Prediction
- SpanEmo: Casting Multi-label Emotion Classification as Span-prediction
- GoEmotions: A Dataset of Fine-Grained Emotions
- SemEval-2018 Task 1: Affect in Tweets