Denoised NER and Relation Extraction Models for Knowledge Graph Population

The patent combines entity extraction and relation extraction into a unified neural network model that processes text and generates both entity mentions and relation triples simultaneously. This integration eliminates the need for separate post-processing steps to normalize relation representations, directly resolving the contradiction between extraction completeness and processing complexity.

The model performs relation extraction in advance during the initial text processing stage, generating relation triples alongside entity mentions in a single pass. This preliminary action prevents the accumulation of complex post-processing requirements that would otherwise be needed to handle relation data separately.

The patent implements an active learning loop where the neural network model processes noisy distant supervision data, identifies uncertain predictions, and queries users for correction only on those specific cases. This feedback mechanism allows the system to leverage efficient distant supervision while maintaining high accuracy by selectively correcting only the noisy portions through user interaction.

Instead of requiring complete manual labeling of all data or accepting all distant supervision data as-is, the system applies partial manual correction only where needed (on uncertain predictions). This selective approach achieves high accuracy without the excessive cost of full manual labeling, resolving the contradiction between labeling efficiency and accuracy.

The system performs manual correction only partially, specifically on data points where the model demonstrates low confidence or high uncertainty. This selective correction approach achieves high overall data accuracy while avoiding the time-consuming and costly process of manual verification for all labeled data, thus resolving the contradiction between accuracy and time investment.

The model serves itself by automatically identifying which data points require human correction based on its own confidence assessments. This self-service capability reduces the need for extensive manual intervention, allowing the system to maintain high accuracy while minimizing time loss.

The patent merges entity extraction and relation extraction capabilities into a single neural network model with unified architecture and loss function. This combination maintains implementation simplicity by using one model instead of separate systems, while simultaneously providing versatile extraction capabilities that handle both entities and relations in a single processing pass.

The neural network model is designed as a universal system that performs multiple functions: it extracts entity mentions, identifies relation types, and generates relation triples all within the same architecture. This multi-functionality resolves the contradiction by providing comprehensive extraction capabilities without requiring separate specialized models for each task.