Neural Ontology Mapping via Word2vec and Doc2vec Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional ontology mapping methods across different domains in biomedical sciences rely heavily on bridging ontologies, which require extensive manual curation and are inefficient for automatic mapping without these intermediaries.
Innovation Solution
A neural document embedding based ontology mapping method using trained word2vec and Doc2vec models to compute vectors for Human Phenotype and Mammalian Phenotype IDs, calculating cosine similarity scores for inter-ontological connections, and sorting results to display relevant mappings without the need for bridging ontologies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If bridging ontologies are used for ontology mapping, then mapping accuracy is improved, but manual curation effort and time increase significantly
Solution Approach 1:
The patent replaces manual mechanical curation processes with automated neural network-based embedding systems. The word2vec and Doc2vec models automatically compute semantic similarities between ontology terms, eliminating the need for manual bridging ontology creation while maintaining mapping accuracy through learned semantic representations.
Solution Approach 2:
The system enables self-service ontology mapping by allowing the neural networks to automatically learn and establish mappings between source and target ontologies without human intervention. The models self-adjust their parameters during training to optimize mapping accuracy, and the system automatically retrieves and processes training data from available resources.
2Adaptability or versatility
If bridging ontologies are created with defined axioms, then inter-ontological connections are established, but system complexity and manual effort increase
Solution Approach 1:
The patent extracts the essential mapping function from complex bridging ontology structures. Instead of maintaining entire bridging ontologies with their axioms and relationships, the system extracts only the necessary mapping relationships through neural network learning, storing compact vector representations that capture semantic relationships without the overhead of full ontology structures.
Solution Approach 2:
The system changes the representation parameters from complex ontology axioms and hierarchical relationships to dense vector embeddings. This parameter transformation allows the system to capture inter-ontological connections in a compact, continuous space that is easier to manipulate and query, reducing system complexity while maintaining adaptability.
3Extent of automation
If conventional mapping methods are used, then ontology terms are mapped across domains, but automation level remains low requiring extensive manual work
Solution Approach 1:
The patent implements feedback mechanisms where the neural networks continuously learn from mapping results and adjust their parameters accordingly. The system uses available gold-standard mappings and annotated data as feedback signals during training, allowing the models to improve their mapping consistency through iterative optimization while increasing automation levels.
Solution Approach 2:
The system performs preliminary actions by pre-training the neural networks on large corpora and available ontology data before actual mapping tasks. This preliminary training establishes consistent parameter settings and learned representations that ensure reliable and consistent mapping results when the automated system is deployed, reducing the need for manual intervention.
Data Source
AI summary
The present disclosure provides a neural document embedding based ontology mapping. Conventional methods that map ontology concepts across domains/species extensively take help of bridging ontologies. Initially the system receives a Human Phenotype (HP) Identification number (ID) pertaining to a phenotype. A first HP ID vector is computed from the HP ID using a trained word2vec model. A second HP ID vector is computed from the HP ID using a trained Doc2vec model. An average HP ID vector is computed based on the first HP ID vector and the second HP ID vector. A plurality of cosine similarity scores are computed based on a comparison between the average HP ID vector and a plurality of average MP ID vectors. The plurality of MP IDs are sorted based on the plurality of cosine similarity scores. The plurality of MP IDs corresponding to the HP ID are selected based on a selection threshold.