Global Local-aware Denoising Framework for Commonsense Knowledge Graph Noise Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise detection methods for Commonsense Knowledge Graphs (CSKGs) struggle to effectively identify noise due to the unique characteristics of CSKGs, including non-canonicalized free-form text nodes and sparse structural information, which are not adequately addressed by existing learning-based and rule-based approaches.
Innovation Solution
The proposed Global and Local-aware Denoising (GOLD) framework uses a PLM-based triple encoder to extract semantic information from CSKGs, combined with a global noise detector that employs rule mining and a local noise detector that utilizes graph neural networks to aggregate neighboring triple information, thereby detecting noise at both global and local levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning-based noise detection methods are applied to CSKGs, then structural relationship capture is improved, but adaptability to non-canonicalized free-form text nodes deteriorates
Solution Approach 1:
The patent segments the noise detection task into two independent detectors: a global noise detector that handles structural relationships and a local noise detector that handles free-form text node semantics. This segmentation allows each detector to specialize in its respective domain, resolving the contradiction between structural capture and text adaptability.
Solution Approach 2:
The patent introduces a rule encoder as an intermediary component that bridges the global and local detectors. The rule encoder processes high-frequency patterns from rule mining and provides guidance to both detectors, enabling them to work together effectively on CSKGs with non-canonicalized nodes.
2Reliability
If rule-based noise detection methods are applied to CSKGs, then logical rule detection is improved, but performance deteriorates due to sparse edge information
Solution Approach 1:
The patent performs preliminary rule mining to extract high-frequency patterns before the actual noise detection. This preliminary action creates a rule bank that guides both the global and local detectors, enabling effective noise detection even when the CSKG has sparse edges, as the rules are pre-computed from available structural information.
Solution Approach 2:
The patent uses a rule encoder that copies and generalizes high-frequency patterns from the mined rules. Instead of learning new rules during detection, the rule encoder replicates and adapts existing patterns, significantly reducing the need for large quantities of training data while maintaining detection effectiveness.
3Reliability
If PLM-based edge classification is applied to CSKGs, then semantic information capture is improved, but computational cost and parameter optimization complexity increase
Solution Approach 1:
The patent segments the PLM-based classification task into a global detector that handles semantic patterns at the triple level and a local detector that handles node-level semantics. This segmentation reduces the parameter optimization complexity by dividing the large optimization problem into smaller, more manageable sub-problems that can be solved independently.
Solution Approach 2:
The patent applies different processing qualities to different parts of the system: the global detector uses rule mining and pattern matching (less computationally intensive) while the local detector uses graph neural networks with aggregated semantic information. This local quality approach optimizes computational resources by applying complex methods only where necessary.
Data Source
AI summary
A system using global and local-aware denoising framework for commonsense knowledge graph noise detection is provided. The system includes an input receiver, a triple encoder, a global noise detector, a local noise detector, and a comprehensive evaluation scorer. The triple encoder receives the CSKG information from the input receiver and extract semantic information from free-text formatted nodes in the CSKG information, so as to obtain structured triples. The global noise detector identifies high-frequency patterns through rule mining and uses a rule encoder to generalize learned rules and guide noise detection. The local noise detector utilizes the structured triples output by the triple encoder to aggregate neighboring triple information around both the head and tail nodes via graph networks. The comprehensive evaluation scorer provides a comprehensive score based on global noise detection and local noise detection, so as to indicate a noise level for the CSKG information.