Webpage Entity Extraction via Bidirectional Text-Structure Iteration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing top-down Hierarchical Conditional Random Fields (HCRF) and Semi-Markov Conditional Random Fields (Semi-CRF) models struggle to accurately identify business names on webpages, as they rely solely on visual layout features, which are insufficient for labeling short isolated strings like business names.
Innovation Solution
A bidirectional, iterative framework is introduced that integrates structure understanding and text understanding components, where the HCRF model uses text-related data from the Semi-CRF model and vice versa, iteratively refining their decisions until a similarity stop criterion is met, leveraging multiple mentions of text across the webpage to enhance labeling accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a top-down HCRF model is used to determine structure first, then a Semi-CRF model is used for labeling, then the labeling process is simplified and faster, but the accuracy of identifying business names and short isolated strings deteriorates
Solution Approach 1:
The patent inverts the traditional top-down approach by implementing a bidirectional framework where text understanding and structure understanding components iteratively refine each other's decisions. Instead of structure determining text labeling in a single pass, the system allows text labeling results to feed back into structure understanding, and structure updates to refine text labeling in subsequent iterations, thereby improving business name identification accuracy while maintaining processing efficiency
Solution Approach 2:
The patent introduces feedback mechanisms where the Semi-CRF model's text labeling results are fed back to the HCRF structure understanding component, and the HCRF's structure decisions are fed back to the Semi-CRF model. This iterative feedback loop allows both components to continuously improve their decisions based on the other's output, resolving the contradiction between processing efficiency and labeling accuracy
2Device complexity
If only visual layout features are used for labeling, then the model complexity is reduced and processing is faster, but the ability to accurately label short isolated strings like business names deteriorates
Solution Approach 1:
The patent merges visual layout features from the HCRF structure understanding component with text-based features from the Semi-CRF model. By combining these two feature types within an iterative bidirectional framework, the system achieves accurate labeling of short isolated strings like business names without excessively increasing model complexity, as the features are integrated through the existing conditional random fields architecture
3Measurement precision
If an iterative bidirectional framework is implemented to improve labeling accuracy, then business name identification improves, but the processing time and computational complexity increase
Solution Approach 1:
The patent applies partial iteration by running the bidirectional framework for a limited number of iterations or until convergence criteria are met, rather than performing infinite iterations. This allows the system to achieve sufficient labeling accuracy for business names and entities while controlling processing time, as the iterative refinements diminish in value after a certain point
Data Source
AI summary
Described is a technology for understanding entities of a webpage, e.g., to label the entities on the webpage. An iterative and bidirectional framework processes a webpage, including a text understanding component (e.g., extended Semi-CRF model) that provides text segmentation features to a structure understanding component (e.g., extended HCRF model). The structure understanding component uses the text segmentation features and visual layout features of the webpage to identify a structure (e.g., labeled block). The text understanding component in turn uses the labeled block to further understand the text. The process continues iteratively until a similarity criterion is met, at which time the entities may be labeled. Also described is the use of multiple mentions of a set of text in the webpage to help in labeling an entity.