Filtering Spurious Knowledge Graph Relationships
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems constructing domain databases or taxonomies include spurious relationships from external knowledge graphs without evaluating their relevance, leading to databases or taxonomies that indicate irrelevant or inapplicable relationships between terms in the input corpus.
Innovation Solution
A system that generates context-based embeddings of terms in the input corpus using methods like Word2Vec or BERT, maps them to vector spaces, and filters out relationships based on similarity thresholds, such as cosine similarity, to determine the relevance of relationships retrieved from knowledge graphs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If relationships are retrieved from external knowledge graphs without evaluation, then the construction process is simple and fast, but the precision and relevance of the constructed domain taxonomy deteriorate due to inclusion of spurious relationships
Solution Approach 1:
The system performs preliminary evaluation of relationships by computing similarity values between term embeddings before finalizing the domain taxonomy construction. This preliminary action filters out spurious relationships early in the process, preventing them from contaminating the final taxonomy while maintaining efficient construction.
Solution Approach 2:
The patent introduces an intermediary evaluation mechanism that computes similarity values between term embeddings as a mediator between the raw knowledge graph relationships and the final domain taxonomy. This intermediary step assesses relationship relevance without requiring complete reconstruction of the taxonomy, balancing precision and productivity.
2Quantity of substance
If all relationships from knowledge graphs are included in the domain taxonomy, then completeness is improved, but noise and irrelevant information increase
Solution Approach 1:
The system changes the parameter of relationship inclusion by introducing a similarity threshold parameter. Relationships are included or excluded based on whether their computed similarity values meet this threshold, transforming the binary include/exclude decision into a parameter-driven filtering process that maintains quantity while improving quality.
Solution Approach 2:
The patent extracts and removes spurious relationships from the knowledge graph relationships before incorporating them into the domain taxonomy. By taking out relationships below the similarity threshold, the system maintains the quantity of relevant relationships while eliminating noise and irrelevant information.
3Measurement precision
If context-based embeddings and similarity computations are performed, then relationship filtering accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The system performs partial similarity computations only for relationships that need evaluation rather than computing all possible similarities. By applying filtering only where necessary and using efficient similarity metrics, the system achieves high filtering accuracy without excessive computational complexity.
Solution Approach 2:
The patent replaces complex manual or rule-based relationship evaluation mechanisms with automated context-based embedding and similarity computation. This substitution uses efficient vector space operations to achieve high filtering accuracy with manageable computational complexity, replacing tedious manual verification processes.
Data Source
AI summary
Systems and techniques that facilitate spurious relationship filtration from external knowledge graphs based on distributional semantics of an input corpus are provided. In one or more embodiments, a context component can generate a context-based word embedding of one or more first terms in a document collection. The embedding can yield vector representations of the one or more first terms. The one or more first terms can correspond to knowledge terms in one or more first nodes of a knowledge graph. In one or more embodiments, a filtering component can filter out a relationship between the one or more first nodes and a second node of the knowledge graph based on a similarity value being less than a threshold. The similarity value can be a function of the vector representations of the one or more first terms. In various embodiments, cosine similarity can be used to compute the similarity value.