Few-Shot Relation Classification With Double-Layer Semantic Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing few-shot relation classification methods face challenges due to the scarcity of annotation instances, especially for emerging relations, leading to noise introduction and limited accuracy, as they primarily focus on local sentence-level characteristics without considering global conceptual semantics.
Innovation Solution
A method involving a coarse-grained and fine-grained double-layer filter system, combined with a false positive instance correction module, to enhance the quality and quantity of training data by identifying high-quality instances with similar entities and relation concepts, and dynamically adjusting the negative sample set proportion during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If distant supervision with strong relation hypothesis is used to increase quantity of instances, then quantity of training data is improved, but noise is introduced to the model
Solution Approach 1:
The patent introduces a neural snowball algorithm as an intermediary mechanism between distant supervision and the classification model. This algorithm uses a Relation Siamese Network (RSN) to measure distances between query instances and supporting seed instances, acting as a filter to identify high-quality positive instances while excluding noisy data, thus resolving the contradiction between increasing data quantity and avoiding noise contamination
Solution Approach 2:
The patent applies local quality by differentiating between different qualities of training instances. Instead of treating all instances uniformly, it uses RSN to evaluate individual instance quality based on semantic similarity to seed instances, allowing the model to selectively utilize high-quality instances while rejecting low-quality or noisy ones, thereby maintaining data quality while expanding quantity
2Measurement precision
If prompt learning focuses on local sentence characteristics to improve instance quality, then local representation is improved, but global conceptual semantics are ignored causing semantic deviation
Solution Approach 1:
The patent merges local sentence-level characteristics with global conceptual semantics by combining prompt learning with knowledge graph information. It integrates entity alignment results, relation types, and conceptual semantics from knowledge graphs with local sentence representations, creating a comprehensive instance representation that captures both local precision and global semantic context, thus preventing semantic deviation
3Device complexity
If excessive dependence on single piece of information is used to simplify processing, then processing complexity is reduced, but semantic deviation occurs hindering accurate understanding
Solution Approach 1:
The patent employs a composite information structure that integrates multiple types of information: local sentence characteristics, entity alignment results, relation types, and global conceptual semantics from knowledge graphs. This composite approach combines diverse information sources into a unified representation framework, maintaining comprehensive semantic understanding while managing processing complexity through structured integration
Data Source
AI summary
A method and an apparatus for few-shot relation classification and filtering, and a device are provided. The method includes: constructing a coarse-grained filter for filtering an unlabeled corpus to obtain candidate instances with a same entity as a seed instance and with similar semantics to the seed instance; constructing a fine-grained filter for filtering the candidate instances to obtain a candidate instance with a same relation concept as the seed instance; defining the candidate instance as a positive instance set, and defining candidate instances with different relation concepts from the seed instance as a negative sample set; constructing a false positive instance correction module for adjusting and controlling a proportion of the negative sample set used by a classifier during training; training the classifier based on a small number of obtained labeled instances belonging to a newly emerging relation and the adjusted positive instance set and negative sample set.