TL;DR
This paper introduces a novel "Explainable Semantic Communication" (TESC) framework for text-based tasks. TESC represents text semantics using interpretable "triplets" (entity-relation-entity) and filters them based on the specific task, significantly reducing transmission data while maintaining or improving performance on tasks like sentiment analysis and question answering.
What Problem Does This Paper Address?
The rapid growth of AI and the Internet of Things (IoT) is leading to an exponential increase in wireless data traffic, exacerbating spectrum scarcity. Traditional communication methods focus on transmitting raw data accurately, which is inefficient for task-oriented applications. Semantic communication aims to transmit only the meaning relevant to a task, improving efficiency. However, existing semantic communication methods often use "black box" neural network outputs (vectors) that are unexplainable and lack the logical structure of human language. This lack of explainability hinders trust and practical adoption. Furthermore, these methods often fail to optimize for specific downstream tasks, leading to unnecessary semantic redundancy and suboptimal compression.
Methods
The proposed TESC scheme addresses these challenges through a two-pronged approach:
- Explainable Semantic Representation:
- Semantic Extraction: Text is converted into triplets (entity-relation-entity). This involves:
- An out-of-the-box triplet extraction tool (like OpenIE).
- A complementary triplet extraction method that uses syntactic dependency analysis to capture more nuanced and implicit relationships (e.g., "bird is red" in addition to "bird flies in sky"). This method leverages tools like Spacy for dependency parsing.
- Semantic Filtering: Extracted triplets are filtered to remove redundancy and task-irrelevant information. This is a two-step process:
- Unique-Relationship Filter: Removes triplets with the same entity pair, ensuring semantic uniqueness.
- Task-Specific Filter: Uses task-specific knowledge (from a knowledge base) to prioritize triplets relevant to the downstream task (e.g., sentiment analysis or question answering). This filter assigns a score to triplets based on their relevance to predefined entity types for the task.
- Task-Oriented Semantic Codec:
- The filtered triplets are encoded by a semantic encoder (e.g., Transformer-based for sentiment analysis, embedding layer for QA) and then transmitted over a wireless channel.
- The semantic decoder at the receiver uses the received semantics to directly perform the downstream task (e.g., MLP for sentiment analysis, Self-Attentive Associative Memory (STM) for QA).
Datasets
- Sentiment Analysis: SST2 dataset (predicting sentiment from movie reviews).
- Question Answering (QA): bAbI dataset (consisting of 20 subtasks, each with 1000 training and 1000 testing questions).
Key Insights and Ideas
- Triplets as Interpretable Semantics: Representing text as triplets provides a structured and human-understandable way to convey meaning, overcoming the "black box" problem of vector-based representations.
- Task-Specific Filtering is Crucial: Discarding task-irrelevant triplets is key to achieving high compression ratios and improving efficiency, as different tasks require different semantic information.
- Dual Semantic Extraction Enhances Completeness: Combining off-the-shelf tools with syntactic dependency analysis allows for a more comprehensive capture of text semantics, including implicit relationships.
- TESC's Versatility: The framework is adaptable to different tasks by customizing the semantic codec and filtering mechanisms, demonstrating its broad applicability.
- Significant Compression Gains: The semantic filtering approach drastically reduces the number of transmitted symbols, especially for complex tasks like QA.
Key Results
- Performance Improvement: TESC shows significant performance gains over traditional communication methods and other semantic communication approaches.
- For sentiment analysis, TESC achieves an 80.5% accuracy improvement over traditional methods.
- For QA, TESC achieves a 150% performance gain over traditional methods.
- Transmission Efficiency:
- In QA tasks, TESC transmits only 8% of the symbols compared to traditional methods.
- The semantic filtering method reduces the average number of triplets by 76.1% and the total word count by 83.8% with minimal performance decay (0.4%).
- Robustness: TESC consistently outperforms baselines like DeepJSCC and DeepSC across various Signal-to-Noise Ratio (SNR) levels and channel conditions (AWGN and Rayleigh fading).
- Computational Efficiency: TESC demonstrates superior computational efficiency, requiring less FLOPS, parameters, and storage space compared to DeepJSCC and DeepSC, making it suitable for resource-constrained devices.
Limitations and Future Work
- Limitations:
- While TESC achieves significant compression, the paper notes that traditional methods can achieve optimal performance at very high SNR levels.
- The current work focuses on text tasks; extending it to other modalities is a natural next step.
- Future Work:
- Exploring explainable semantic communication for computer vision tasks.
- Developing a unified semantic representation method for multimodal data.
Summary updated