Knowledge Graph Entity Classification via Adversarial Embedding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining knowledge graphs struggle to efficiently combine different types of embeddings in a shared space without forming embedding-type clusters, which hinders accurate entity identification and classification.
Innovation Solution
A device and method that utilize a classification decision model to determine whether a word is an entity or not, and to which type of entity it belongs, by generating second embeddings using a generator trained with adversarial learning and gradient reversal techniques, effectively reducing the number of parameters required for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple embeddings are combined using concatenation, then the classification model can capture diverse features, but the number of parameters increases significantly
Solution Approach 1:
The patent merges multiple embedding types (word embeddings, character embeddings, domain-specific embeddings) into a unified shared space through a generator function, allowing the model to leverage diverse features while maintaining a compact parameter count. The generator maps different embedding types to a common representation space, enabling effective combination without concatenation's parameter explosion.
Solution Approach 2:
The patent transforms the parameter structure by using a generator function with shared parameters that maps multiple embedding types to a unified space. Instead of maintaining separate parameters for each embedding type through concatenation, the generator uses parameter sharing and transformation to achieve efficient representation with reduced parameter count.
2Adaptability or versatility
If different types of embeddings are mapped to a shared space, then the model achieves better generalization, but embedding-type clusters may form which reduce classification accuracy
Solution Approach 1:
The patent introduces a discriminator as an intermediary component that detects and prevents the formation of embedding-type clusters in the shared space. The discriminator acts as a mediator between the generator and the classifier, ensuring that embeddings from different types are properly mixed and not segregated into distinct clusters, thereby maintaining classification accuracy.
Solution Approach 2:
The patent applies preliminary anti-action by using adversarial training where the discriminator is trained to identify embedding-type clusters before they can negatively impact classification. The generator is trained to counteract the discriminator's detection, creating a preliminary defense against cluster formation that preserves classification performance.
3Measurement precision
If adversarial learning with gradient reversal is used, then embedding-type clusters are prevented, but the training process becomes more complex
Solution Approach 1:
The patent implements feedback through the adversarial training loop where the discriminator's gradient is reversed and fed back to the generator. This feedback mechanism allows the generator to learn how to prevent embedding-type cluster formation by observing the discriminator's predictions and adjusting its mapping accordingly, improving classification accuracy through iterative refinement.
Data Source
AI summary
A device and method for determining a knowledge graph. A second embedding is determined for a first embedding for a word including a function. A first classification, which determines whether or not the word is an entity for the knowledge graph, or which defines to which entity or to which type of entity for the knowledge graph the word in the knowledge graph is to be assigned, is determined for the second embedding using a first classifier. A second classification, which defines to which type of embeddings from a plurality of types of embeddings the second embedding is to be assigned, is determined for the second embedding using a second classifier. At least one parameter for the function is trained in a training as a function of a gradient for the training of the first classifier and as a function of a gradient for the training of the second classifier.