Introduction
In the modernization of power distribution networks, the transition from manual dispatching to "intelligent virtual dispatchers" is a critical step. These systems must accurately interpret voice-recognized reports from field personnel and map them to a structured Knowledge Graph (KG). However, real-world utility operations present significant challenges for standard Entity Linking (EL) processes.
Technical terminology in power grids is often complex and prone to transcription errors during speech-to-text conversion. For instance, regional accents or environmental noise can cause "Qianjiang 16730 switch" to be misidentified. Furthermore, technical synonyms (e.g., "knife switch" vs. "knife gate") and the presence of multiple entities in a single command (e.g., "Line A and Line B") complicate simple string-matching or purely semantic models. This paper proposes a specialized approach to overcome these hurdles by integrating phonetic and lexical features into a deep learning framework.
Methodology: The LSF-SCNN Model
The researchers developed an improved Lexical Semantic Feature based Skip Convolution Neural Network (LSF-SCNN). The core innovation of this model is its multi-modal representation of text, which moves beyond simple word meanings to include how words sound and their grammatical roles.
Multi-Dimensional Feature Matrices
Instead of a single vector, each word is represented by a three-layer matrix:
- Semantic Layer: Utilizes word2vec (Skip-gram) to capture the underlying meaning of terms.
- Phonetic Layer: Employs a custom pinyin2vec mapping. This allows the model to recognize that two words might be the same entity if they sound similar, even if their transcribed spellings differ due to speech recognition errors.
- Lexical Layer: Captures Part-of-Speech (POS) features, leveraging the fact that power grid commands often follow rigid structural patterns like "Location + ID + Device."
Architecture and Attention
The model uses wide convolution and k-max average pooling (KMA) to generate representation vectors for both the input text and the KG entities. An integrated attention mechanism dynamically weights these features, allowing the model to prioritize phonetic similarity when transcriptions are garbled or semantic meaning when the text is clear.
Experimental Evaluation
The study utilized a robust dataset of 40,000 dispatching texts from the Hangzhou Power Supply Company, containing approximately 112,000 positive sample pairs for training. The proposed model was benchmarked against seven control models, ranging from basic string matching to standard deep learning architectures.
Key Results
The LSF-SCNN model significantly outperformed all baselines. As shown in the performance data, the proposed model achieved an accuracy of 91.94%, a substantial improvement over direct matching (59.17%) and the standard LSF-SCNN without the new feature dimensions (72.09%).
Efficiency is also a key factor for real-time dispatching. Despite its multi-layered complexity, the model maintains an average processing time of 0.471 seconds per text, making it suitable for live deployment in power distribution centers.
Conclusion
By incorporating phonetic and lexical layers, the LSF-SCNN model provides a robust solution for entity linking in noisy, technical environments. While the 91.94% accuracy is a major advancement, future work will focus on handling highly ambiguous commands and rare vocabulary. The authors suggest implementing a "human-in-the-loop" system where the virtual dispatcher can request clarification when confidence scores are low, ensuring the near-perfect reliability required for power grid operations.
Summary updated