Concept Extraction Denoising via Vector Embedding Similarity

The patent introduces vector representations as an intermediary between surface-level word sequences and concept annotations. Instead of directly mapping words to concepts using simple algorithms, the system transforms words into vector embeddings that capture semantic relationships, then uses these vectors to compute similarity with reference concepts. This intermediary representation enables contextual understanding while maintaining computational efficiency.

The patent changes the parameter representation from discrete surface forms (words) to continuous vector embeddings. By transforming the input representation into a different parameter space where semantic similarity is captured through vector distance metrics, the system achieves both accuracy and efficiency. The vector space allows for nuanced comparisons that go beyond exact word matching.

The patent uses pre-trained vector embeddings that capture linguistic knowledge without requiring domain-specific training data. Instead of training custom classifiers for each domain, the system copies proven vector representation techniques from natural language processing research and adapts them to concept extraction. This approach achieves high accuracy without the complexity of domain-specific supervised learning pipelines.

The vector embedding approach serves multiple functions simultaneously: it captures semantic meaning, enables similarity computation, and provides a unified representation for both words and concepts. This universal representation eliminates the need for separate processing pipelines for different annotation tasks, reducing overall system complexity while maintaining discrimination accuracy.