Webpage Entity Recognition via Graph Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current natural language processing models can only extract entities from webpages but cannot determine associations between them, limiting their effectiveness in information extraction tasks.
Innovation Solution
An entity recognition method that extracts specified entities from webpages, encodes their text and layout information, and uses graph convolution to recognize relationships between entities, considering both content and layout for accurate association detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional natural language processing models are used to extract entities from webpages, then entity extraction can be performed, but association detection between entities is not possible
Solution Approach 1:
The patent combines traditional NLP entity extraction with graph convolutional networks to create a unified model that performs both entity extraction and association detection. The text encoding results and position encoding results are merged into triples that serve as input to the graph convolution model, enabling simultaneous achievement of both capabilities.
Solution Approach 2:
The patent transitions from traditional 1D text processing to 2D/3D spatial reasoning by incorporating position encoding that captures the spatial relationships between entities on the webpage. The graph convolution operates on triples that include positional information, adding a spatial dimension to the entity extraction process.
2Device complexity
If only text content is used for entity recognition, then processing is simple, but association detection accuracy is limited
Solution Approach 1:
The patent applies different encoding strategies to different aspects of the data: text encoding for content and position encoding for spatial relationships. Each encoding method is optimized for its specific purpose, with text encoding capturing semantic information and position encoding capturing spatial arrangement, thereby improving overall association detection accuracy.
Solution Approach 2:
The patent creates a composite representation by combining text encoding results and position encoding results into triples. This composite structure integrates multiple types of information (semantic and spatial) that work together to improve association detection beyond what either encoding could achieve alone.
3Measurement precision
If graph convolution is applied to recognize relationships between entities, then association detection accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the computational process into distinct stages: text encoding, position encoding, triple construction, and graph convolution. This segmentation allows each component to be optimized independently and enables parallel processing of multiple triples, reducing overall computational complexity while maintaining association detection accuracy.
Data Source
AI summary
An entity recognition method, a model training method, an electronic device, and a medium, which relate to fields of artificial intelligence, information acquiring technologies. The entity recognition method includes: extracting specified entities from a text in a source file of a webpage to be recognized, and acquiring a text encoding result for each specified entity; determining a text block formed by each specified entity in the webpage, and encoding a relative layout information between each two text blocks, to obtain a position encoding result; constructing a triple by the position encoding result for each two text blocks and the text encoding results for respective specified entities of the two text blocks; and performing a graph convolution on each triple to obtain a relation recognition result for the webpage to be recognized, where the relation recognition result indicates whether an association exists between each two text blocks in the webpage.