Chinese Concept Embedding via Wikipedia Link Structure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for learning word representation in natural language processing fail to distinguish between different concepts of a word, particularly in the Chinese language, due to their reliance on word form and inability to handle polysemy effectively.
Innovation Solution
A method and device for Chinese concept embedding generation based on Wikipedia link structure, which constructs a link information database, builds a training dataset, and trains a concept embedding model using positive and negative training instances to generate accurate concept embeddings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional word embedding methods are used to learn word representation, then the embedding can be learned efficiently through co-occurrence information, but the method fails to distinguish different concepts/meanings of a word (polysemy problem)
Solution Approach 1:
The patent segments the unified word embedding into multiple concept-specific embeddings. By using Wikipedia link structures, the method identifies different concepts associated with a word and generates separate embedding representations for each concept, thereby resolving the polysemy problem while maintaining computational efficiency through the structured approach.
2Loss of information
If word form-based distinction is used, then the method is simple and efficient, but it cannot distinguish different meanings corresponding to the same word form
Solution Approach 1:
The patent introduces Wikipedia link structures as an intermediary to bridge the gap between word forms and their multiple meanings. The link structures serve as a mediator that connects a word to its different concepts, allowing the system to retrieve concept-specific information without directly analyzing complex semantic relationships, thus preserving information while maintaining efficiency.
3Reliability
If unified embedding representation is learned for a word, then the learning process is straightforward, but all concepts of the word are confused together without accurate distinction
Solution Approach 1:
The patent adds a new dimension to the embedding space by incorporating concept identifiers from Wikipedia link structures. Instead of a single embedding vector for a word, the system creates multiple embedding vectors organized in a hierarchical structure where each concept has its own embedding, enabling accurate concept distinction while building upon the familiar embedding framework.
Data Source
AI summary
A method and a device for Chinese concept embedding generation based on Wikipedia link structure includes: Step (1): According to the title concepts and/or link concepts in Chinese Wikipedia pages, a link information database is constructed; Step (2): For the title concepts, according to their link relationships with link concepts in the link information database, the positive and negative training instances are constructed respectively, which constitute the training dataset; Step (3): A concept embedding model is built, including an input layer, an embedding layer, a computational operation layer, and an output layer; Step (4): The concept embedding model is trained with the training dataset, then, the Chinese concept embedding is extracted/generated from the concept embedding model. The method can accurately distinguish different concepts and overcome the problem of polysemy that troubles the traditional embedding methods, which is beneficial to generate more accurate concept embedding representation.