Neural Entity Linking via Dual-Path CNN-LSTM Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional entity linking techniques are inadequate for organizations as they rely on Wikipedia data, which lacks coverage of financially relevant entities and context information, making it difficult to link mentions in textual sources to entities in knowledge graphs effectively.
Innovation Solution
An end-to-end neural entity linking module is developed, utilizing a machine learning model to compute entity embeddings with a margin-based loss function and deploy a wide and deep learning model to match character and semantic information, independent of Wikipedia, enabling effective linking of named mentions to corresponding entities in knowledge graphs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional entity linking techniques using Wikipedia data are employed, then the system achieves good performance on Wikipedia-based entities, but it fails to generalize to enterprise-specific entities that are not covered in Wikipedia
Solution Approach 1:
The patent segments the entity linking task into two independent components: character-level matching (using CNN) and semantic-level matching (using LSTM). This segmentation allows the system to process entities without relying on Wikipedia context, improving adaptability to enterprise-specific entities while maintaining linking accuracy through specialized processing for each aspect.
Solution Approach 2:
The patent changes the fundamental parameters of the entity linking approach by transitioning from context-dependent Wikipedia-based models to context-independent enterprise-specific models. It uses margin-based loss functions and specialized neural network architectures (CNN for character patterns, LSTM for semantic patterns) that do not require Wikipedia training data, thereby expanding entity coverage while maintaining reliability.
2Reliability
If pre-trained models leveraging Wikipedia context data are used, then the models achieve great performance, but there is insufficient information for re-training or fine-tuning when working with internal organizational data
Solution Approach 1:
The patent implements self-service by designing a model that generates its own entity embeddings from scratch using margin-based loss functions, without requiring pre-trained Wikipedia context data. The system serves itself by learning character and semantic patterns directly from enterprise data, eliminating the dependency on external Wikipedia resources while maintaining high performance.
Solution Approach 2:
The patent performs preliminary action by pre-defining the dual-path architecture (CNN for character patterns, LSTM for semantic patterns) and margin-based loss functions before training. This preliminary setup enables the model to be trained directly on enterprise data without requiring prior Wikipedia-based pre-training, thus improving data source flexibility while ensuring reliable performance through carefully designed preprocessing and architecture.
3Adaptability or versatility
If conventional entity linking methods based on standard Wikipedia datasets are applied, then the methods work well for Wikipedia entities, but they cannot handle entities of financial interest that are not present in Wikipedia
Solution Approach 1:
The patent introduces another dimension by adding character-level pattern matching (CNN) as a complementary dimension to semantic-level matching (LSTM). This dual-dimensional approach allows the system to capture entity information from multiple perspectives, improving coverage of diverse entity types including financial entities not in Wikipedia, while compensating for the loss of contextual information through character-level features.
Data Source
AI summary
Various methods, apparatuses/systems, and media for end-to-end entity linking are disclosed. The system includes a processor; and a memory operatively connected to the processor via a communication interface, the memory storing computer readable instructions, when executed, causes the processor to: detect all named entity mentions from a plurality of data sources; compute, in response to detecting, entity embeddings in a knowledge graph by implementing context information and a margin-based loss function; validate the entity embeddings; deploy, in response to validating the entity embeddings, a machine learning model to match character and semantic information, respectively; and link, in response to deployment of the wide and deep learning model, the named mentions in text with corresponding entities in the knowledge graph.