Knowledge Graph Relation Fusion via Prompt-Based Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Knowledge Graph (KG) acquisition faces challenges in efficiently merging relations extracted from different data sources, leading to inefficiencies and redundancies in KGs due to the inability to identify and combine similar relations, resulting in large, inefficient graphs with irrelevant and repeating knowledge instances.
Innovation Solution
A prompt-based learning method is employed to identify and merge relations by generating prompts for language models to predict candidate predicates, clustering these predictions, and fusing relations to generate an optimized KG with reduced repetition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If relations are extracted from multiple data sources without fusion, then the coverage of knowledge instances is improved, but the KG becomes inefficient and redundant with repeating knowledge instances
Solution Approach 1:
The patent applies merging by combining multiple extracted relations that represent the same semantic concept into a single fused relation. The system identifies relations with similar head entities, tail entities, and predicate meanings, then merges them into one canonical relation representation, eliminating duplicates while preserving knowledge coverage from multiple sources.
Solution Approach 2:
The patent creates a universal relation fusion mechanism that handles various types of relations from different data sources through a common framework. The fusion system universally processes relations by comparing their semantic equivalence using string similarity metrics and contextual information, applying the same fusion logic across diverse relation types and sources.
2Loss of information
If all extracted relations are retained without merging, then the completeness of knowledge is improved, but the size of the KG increases with irrelevant and repeating instances
Solution Approach 1:
The patent extracts only the essential and unique relation information by identifying and removing duplicate relations. The fusion process extracts the canonical form of each relation by selecting representative predicates and entities, keeping only one instance of each semantically equivalent relation while discarding redundant copies from different sources.
Solution Approach 2:
The patent merges multiple occurrences of the same relation into a single entry, combining their information while eliminating size inflation. When relations are identified as semantically equivalent through similarity comparison, their knowledge content is merged into one relation entry, preserving completeness without increasing graph size proportionally to the number of sources.
3Productivity
If relation fusion is implemented to reduce redundancy, then the efficiency of the KG is improved, but the complexity of the fusion process increases
Solution Approach 1:
The patent segments the relation fusion process into distinct computational stages: extraction of relation triples from text, comparison of head and tail entities for equivalence, analysis of predicate similarity using string metrics, and final fusion decision-making. This segmentation allows each subtask to be handled by specialized algorithms, improving overall efficiency while managing complexity through modular processing.
Solution Approach 2:
The patent introduces intermediary computational mechanisms including entity embedding representations and predicate similarity scoring functions that mediate between raw extracted relations and final fused results. These intermediaries translate diverse relation formats into a common representation space, enabling efficient comparison and fusion while abstracting away the complexity of direct relation-to-relation matching.
4Measurement precision
If string similarity metrics are used for predicate comparison, then the accuracy of relation fusion is improved, but the computational time increases
Solution Approach 1:
The patent applies partial action by using string similarity metrics selectively rather than exhaustively for all relation comparisons. The system first performs quick filtering using entity equivalence checks and basic predicate matching, then applies more computationally intensive string similarity measurements only to candidate relations that pass the initial filter, reducing overall computational time while maintaining fusion accuracy for relevant cases.
Data Source
AI summary
A method of generating knowledge graphs is disclosed which includes receiving a first input dataset representing unstructured sentences each having a head, a tail, and a predicate, receiving a second input dataset representing classification of heads and tails, generating a first structured database including a plurality of rows each row including a masked sentence where the sentence from the first structured database is substituted with a mask, generating prompts for the masks associated with each row in the first structured database to thereby generate a second structured database; communicating with a language model (LM) the second structured database to thereby generate and process candidate predicates, numerically representing the predicate candidates based on the associated probability numbers, clustering the numerical representation of the predicate candidates, fusing relations of the predicate candidates based on the clustering, and generating an optimized knowledge graph based on reduced repetition of candidate predicates.