Distributed Representation Set Modification via Relationship Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for modifying distributed-representation sets in document data processing primarily focus on learning synonym and antonym pairs, failing to effectively utilize data pairs with various relationships, such as intermediate or distant meaning connections.
Innovation Solution
A document data processing apparatus and method that modifies distributed-representation sets by learning multiple data pairs with corresponding scores, allowing for the inclusion of various relationship types and using a loss function to minimize the difference between the relationship values and target scores, thereby improving the quality of the representation set.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If only synonym pairs and antonym pairs are used for learning in distributed-representation set modification, then the learning process is simple and focused, but data pairs with various relationships (intermediate relationships, close meaning relationships, distant meaning relationships) cannot be effectively utilized
Solution Approach 1:
The patent introduces a relationship type parameter that categorizes data pairs into different relationship types (synonym, antonym, intermediate, close meaning, distant meaning). By changing the parameter space to include multiple relationship types rather than just two, the system can utilize diverse data pairs for learning while maintaining a structured approach through parameterized relationship classification.
2Manufacturing precision
If multiple data pairs with various relationships are subjected to learning, then the quality of distributed-representation set is improved, but the complexity of the learning process increases
Solution Approach 1:
The patent segments the learning process by introducing relationship type as a segmentation parameter. Data pairs are segmented into different groups based on their relationship types (synonym, antonym, intermediate, close meaning, distant meaning). This segmentation allows the system to handle complex multi-type learning by breaking it down into manageable segments, each processed with appropriate weighting, thereby improving representation quality while controlling learning complexity through structured segmentation.
3Measurement precision
If synonym pairs and antonym pairs are subjected to learning, then the loss function can be minimized with focused data, but data pairs having various relationships are not utilized as learning targets
Solution Approach 1:
The patent makes the learning system universal by designing it to handle multiple relationship types through a unified framework. The loss function is generalized to accommodate different relationship types (synonym, antonym, intermediate, close meaning, distant meaning) with appropriate weighting. This multi-functionality allows the same learning mechanism to process diverse data pairs, improving both the precision of relationship representation and the versatility of learning data utilization.
Data Source
AI summary
A document data processing apparatus includes a memory and a processor. The memory stores a distributed-representation set including multiple distributed representations corresponding to multiple pieces of data. The processor is configured to modify the distributed-representation set on the basis of multiple data pairs and multiple scores corresponding to the data pairs. The data pairs are subjected to learning. The processor is configured to modify the distributed-representation set in such a manner that, for each of the data pairs, a value indicating a relationship in a modified distributed-representation pair corresponding to the data pair comes close to a score corresponding to the data pair.