TESC: Explainable Semantic Communication

• Triplet-based Explainable Semantic Communication (TESC) is a task-focused framework using explicit triplet representations to enable efficient, interpretable, and knowledge-aligned semantic transmission over noisy channels.
• It integrates methods like OpenIE and syntactic dependency parsing to extract and merge semantic triplets, enhancing completeness by capturing nuanced linguistic relations.
• Empirical results demonstrate TESC's superior performance in sentiment analysis and question answering through significant symbol count reductions and competitive accuracy under challenging channel conditions.

Triplet-based Explainable Semantic Communication (TESC) is a task-oriented semantic communication scheme for text tasks that achieves efficient, interpretable, and knowledge-aligned information transfer using explicit triplet representations. The TESC framework integrates semantic extraction, knowledge-based filtering, and neural encoding—customized per task—to enable explainable and compact semantic transmission over noisy wireless channels. Empirical results demonstrate superior performance compared to both conventional and deep learning–based semantic communication baselines, particularly in challenging settings with aggressive bandwidth constraints and lossy physical channels (Liu et al., 2023).

1. Formal Problem Setting

TESC is designed for inference-focused communication tasks, such as sentiment analysis (SA) and question answering (QA), where the transmitter observes a text source $I = [i_1, \ldots, i_N]$ and aims to communicate sufficient semantic content to allow the receiver to perform downstream inference. The end-to-end system can be formalized as a pipeline: $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$ The overall training objective follows a rate–distortion framework: $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$ where $L_{\text{task}}$ measures inference error (e.g., cross-entropy), $q(I)$ is the ground-truth, $H(S)$ is the entropy (average transmitted symbols), and $\lambda$ balances task accuracy against transmission efficiency (Liu et al., 2023).

2. Semantic Extraction Pipeline

Semantic extraction in TESC proceeds via two complementary methods to ensure completeness and explainability.

2.1 OpenIE Component:

An Open Information Extraction (OpenIE) annotator is applied to source text $I$ to obtain initial triplets $A_{\text{OpenIE}}$.

2.2 Syntactic Dependency Augmentation:

Using spaCy, a dependency parse tree $A_d = S_d(I)$ is constructed. A rule-based extraction process $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$0 identifies, for each verb $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$1, its syntactic subject $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$2 and object $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$3, expands these entity nodes to include modifiers, and forms additional triplets $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$4 for information potentially missed by OpenIE.

2.3 Deduplication and Merging:

The full set of extracted triplets is given by $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$5, where $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$6 is the output of the syntactic dependency extractor. This combination yields higher semantic completeness, as evidenced by the recovery of modifier relations omitted in OpenIE, e.g., capturing "bird is red" from "a red bird flies in the sky".

3. Knowledge-Based Semantic Filtering

TESC introduces a knowledge-driven, two-stage filtering process to minimize redundancy and focus on task-relevant information.

3.1 Duplicate Removal:

Applying $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$7, only the first triplet for each unique entity pair (head, tail) is retained, producing $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$8.

3.2 Task-Relevant Pruning:

For a given task, further filtering $$\begin{align*} &\text{Semantic extraction: } A = R_T(I),\quad \text{where } A = \{a_1, \ldots, a_K\},\quad a_k = (\text{en}_{k,i}, r_{k,ij}, \text{en}_{k,j}) \ &\text{Semantic filtering: } X = C_T(A),\quad X \subset A\ &\text{Semantic encoding: } M = E_{\alpha_T}(X) \ &\text{Channel encoding: } S = Q_{\sigma}(M) \ &\text{Wireless channel: } Y = hS + n,\quad n \sim \mathcal{CN}(0, \sigma^2 I) \ &\text{Channel decoding: } M' = Q_{\chi}^{-1}(Y) \ &\text{Semantic decoding: } p = E_{\mu_T}^{-1}(M') \end{align*}$$9 is performed using small scoring functions:

• For Sentiment Analysis: A triplet score $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$0 is computed, favoring longer entities/relations or the presence of sentiment-laden adjectives/verbs. Bottom $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$1 are pruned.
• For Question Answering: Triplets are scored by the count of question-relevant entity types $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$2 (e.g., time, location, number): $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$3. Only triplets with $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$4 are retained.

Empirically, average triplet count drops from 70.8 to 16.9 (–76.1%) and average word count from 517.3 to 83.6 (–83.8%) post-filtering, with only a 0.4% accuracy reduction.

Stage	Avg. Triplets	Avg. Words	Accuracy Drop
Pre-filter	70.8	517.3	–
Post-filter	16.9	83.6	0.4%

4. Semantic Encoding and Channel Codec Architectures

TESC's encoder and decoder architectures are tailored per task and quantized for communication efficiency.

4.1 Sentiment Analysis:

• Semantic encoder $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$5: BERT-style Transformer with 12 heads, 768-dim embeddings, 256-dim hidden layer, followed by a 192-unit ReLU dense.
• Semantic decoder $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$6: MLP mapping 192 to 2 outputs, with softmax.

4.2 Question Answering:

• Semantic encoder $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$7: Triplet embedding layer mapping each triplet to a 64-dim vector.
• Semantic decoder: STM (Structured Two-Memory) architecture with item memory $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$8 and relational memory $$\min_{\text{all params}}\, \mathbb{E}_{I,T} [ L_{\text{task}}(p(I), q(I)) ] + \lambda \cdot H(S)$$9, updated via parameterized functions and outputting the final answer vector $L_{\text{task}}$0 at each timestep.

All networks are quantized to 8 bits/weight. The channel model comprises a linear layer combined with AWGN or Rayleigh fading; bit allocation is learned implicitly.

5. Task-Specific Customization and Training

Task-specific customization occurs at both semantic filter and decoder head levels:

• Sentiment Analysis: Filter favors sentiment-rich triplets; output head is a 2-way softmax for polarity classification.
• Question Answering: Pruning is based on matching question types; decoder is STM as above.

Training employs cross-entropy loss: $L_{\text{task}}$1

$L_{\text{task}}$2

Evaluation metrics include Top-1 accuracy vs. channel SNR, mean number of transmitted symbols/sentence, FLOPs, and model footprint.

6. Comparative Experimental Results

Experiments are conducted on SST-2 (sentiment) and bAbi (QA, 20 subtasks), with competing baselines: Error_free, DeepJSCC, DeepSC + classifier, and Huffman+RS+16QAM.

6.1 Accuracy and Efficiency:

• Sentiment Analysis (5 dB AWGN): TESC outperforms Huffman+RS by 80.5% accuracy gain.
• Question Answering (5 dB Rayleigh): TESC achieves +150% over Huffman+RS, +20.6% over DeepSC, +7.5% above DeepJSCC. TESC nearly matches Error_free at SNR ≥ 15 dB.

6.2 Symbols per Sentence:

• Sentiment Analysis: TESC ≈ 45, DeepSC ≈ 38, DeepJSCC ≈ 60, Huffman+RS ≈ 180.
• QA: TESC ≈ 10, DeepSC ≈ 25, DeepJSCC ≈ 30, Huffman+RS ≈ 120. TESC achieves 8% of traditional symbol budget in QA.

6.3 Model Complexity (QA Task):

• | Model | FLOPs | Params | Model Size (8-bit) | |-----------|-----------|---------|--------------------| | TESC | $L_{\text{task}}$3 | 1.20M | 1.14 MB | | DeepJSCC | $L_{\text{task}}$4 | 1.37M | 5.21 MB | | DeepSC | $L_{\text{task}}$5 | 3.33M | 12.3 MB | | Huffman+RS| $L_{\text{task}}$6 | – | – |

Visual QA case studies show TESC enabling correct response without explicit text reconstruction, unlike competing schemes.

7. Significance, Limitations, and Outlook

TESC provides a model-agnostic, interpretable, and knowledge-aligned approach to semantic communication, achieving compression and transmission efficiency while preserving task performance. The explicit use of triplet forms yields inherent explainability—filtered semantic units can be directly audited for information content and task relevance, addressing the explainability gap inherent in prior neural semantic communication schemes (Liu et al., 2023).

The tightly integrated pipeline—combining explicit linguistic structure extraction, knowledge-driven filtering, and deep neural coding—confers unique transmission efficiency. Empirical evidence for both sentiment analysis and question answering indicate near upper-bound performance at moderate SNR, substantial symbol count reductions, and competitive computational cost.

A plausible implication is that TESC's explicit explainable representation framework may generalize to more complex multi-modal or reasoning-centric communication tasks, where semantic transparency and efficient rate usage are co-critical. An outstanding area for further investigation is the impact of larger and more varied knowledge bases for filtering, and the system's adaptability to open-domain or dynamic-task scenarios.