Financial Relation Extraction via Attention and Prior Knowledge
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing relation extraction models perform poorly when applied to Chinese texts and fail to effectively utilize prior knowledge in specific sectors like finance, lacking generalization ability and semantic interpretability, and are inadequate for multi-entity, multi-relation tasks.
Innovation Solution
A relation extraction system that incorporates a deep pretraining module, keyword analyzing module, and attention mechanism module, using Chinese finance-related corpora and prior knowledge to generate text feature vectors and predict financial-entity relations with improved semantic expression and interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If general relation extraction models designed for English texts are applied to Chinese texts or sector-specific tasks, then the model structure can be reused, but the extraction accuracy and semantic interpretability deteriorate
Solution Approach 1:
The patent applies local quality by introducing sector-specific knowledge bases and attention mechanisms that focus on domain-relevant entities and relations. The system dynamically adjusts the extraction process based on the specific sector (finance, healthcare, etc.), applying specialized knowledge only where needed rather than using a uniform approach across all domains. This resolves the contradiction by maintaining model structure reusability while improving accuracy through localized domain adaptation.
Solution Approach 2:
The patent changes key parameters by introducing sector-specific entity types, relation types, and knowledge base configurations. The system dynamically adjusts extraction thresholds, attention weights, and knowledge integration strategies based on the target domain. This allows the same base model to adapt to different sectors with high accuracy by modifying domain-specific parameters rather than the overall model architecture.
2Measurement precision
If deep learning models use large pretrained models and massive corpora to extract text features, then the semantic representation capability is improved, but the generalization ability and semantic interpretability deteriorate
Solution Approach 1:
The patent introduces a knowledge base as an intermediary between the pretrained language model and the relation extraction task. This knowledge base acts as a mediator that translates general semantic representations into domain-specific insights, improving both interpretability and generalization. The attention mechanism serves as another intermediary that selectively focuses on relevant entities and relations, bridging the gap between comprehensive semantic understanding and targeted domain application.
Solution Approach 2:
The patent segments the relation extraction process into distinct modules: entity recognition, relation prediction, and knowledge base integration. Each module can be independently optimized and adjusted for different domains. This segmentation allows the system to maintain strong semantic representation capabilities while improving generalization by allowing domain-specific adjustments in each module without retraining the entire model.
3Productivity
If existing relation extraction models focus only on text feature extraction, then the processing speed is maintained, but the generalization ability and semantic interpretability deteriorate
Solution Approach 1:
The patent performs preliminary action by pre-building sector-specific knowledge bases and pre-computing entity embeddings before the actual relation extraction process. This preprocessing step enables the model to quickly access domain-specific knowledge during inference without significantly increasing processing time. The attention mechanisms are also pre-configured with domain-relevant weights, allowing fast processing while maintaining strong generalization ability through domain-adapted features.
Data Source
AI summary
The present invention relates to a relation extraction system adapted to financial entities and fused with prior knowledge and a method thereof, the system at least comprising: a deep pretraining module, for training and generating a deep pretrained model for recognizing attributes of the financial entities; a keyword analyzing module, for extracting and outputting positional information and importance vectors of keywords in a Chinese finance-related text; an attention mechanism module, for encoding the positional information of the keywords to obtain attention masks, and inputting them with entity information into the deep pretrained model to acquire text feature vectors; and an optimal margin distribution model module, for predicting financial-entity relations based on the text feature vectors and the importance vectors. Aiming at low applicability of existing models to specific Chinese fields, the present invention obtains more accurate extraction results of entities and related features in Chinese finance-related texts.