Semantic Relationship Learning With Similarity-Ranged Negative Examples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semantic relationship learning devices struggle to effectively discriminate between negative examples with similar and dissimilar meanings, leading to suboptimal machine learning performance.
Innovation Solution
A semantic relationship learning device that classifies negative example data pairs into similarity level ranges based on calculated feature values, allowing for a structured selection of learning-purposed negative example data sets to improve the machine learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If negative examples with extremely close meanings and extremely far meanings are mixed together in machine learning training, then the training data coverage is improved, but the semantic relationship discrimination performance deteriorates
Solution Approach 1:
The patent segments negative examples into multiple groups based on similarity levels (first similarity level range for close meanings, second similarity level range for far meanings). This segmentation allows the system to process different types of negative examples separately, preventing the degradation of discrimination performance while maintaining comprehensive training data coverage.
Solution Approach 2:
The patent applies different processing strategies to different parts of the training data. Specifically, it uses different similarity level ranges for different learning phases or gives different weights to different negative example groups, ensuring that each part of the training data contributes optimally to the semantic relationship discrimination task.
2Productivity
If all negative example data pairs are used simultaneously in machine learning training, then the learning efficiency is improved, but the convergence to optimal solution deteriorates
Solution Approach 1:
The patent performs preliminary classification of negative examples into different similarity level ranges before the main learning process. This preliminary action organizes the training data in a way that facilitates more effective learning convergence, allowing the system to first learn from easier distinctions (far meanings) before tackling harder distinctions (close meanings).
Solution Approach 2:
The patent implements a dynamic learning approach where the system can adjust which negative example groups are used at different learning stages. This dynamic selection allows the learning process to adapt to the model's current capability level, improving both efficiency and convergence reliability.
Data Source
AI summary
A semantic relationship learning device includes processing circuitry to acquire positive example data pairs and to generate negative example data pairs by combining the language data, each of the negative example data pairs being formed with language data that are not in the predetermined semantic relationship with each other; to extract feature values from the negative example data pairs; to calculate a similarity level between the feature values; to classify the negative example data pairs into predetermined similarity level ranges or classifies learning-purposed negative example data pairs into predetermined similarity level ranges, thereby generating learning-purposed negative example data sets corresponding to the similarity level ranges; to select a learning-purposed negative example data set from the learning-purposed negative example data sets in an order according to a selection schedule; and to perform a machine learning process by using the selected learning-purposed negative example data set and the positive example data pairs.