Document-Level Event Extraction with Iterative Pre-Filling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Legacy event extraction methods face challenges in accurately extracting event records without trigger words, leading to performance fluctuations and under-fitting due to manual selection of event role generation sequences and parallel generation methods.
Innovation Solution
A document-level event extraction method employing an iteratively parallel generation with pre-filling strategy (IPGPF) that iteratively generates event roles without manual sequencing, using a pre-filling strategy to mitigate under-fitting and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual selection of event role generation sequences is used, then event extraction can be performed, but performance fluctuations occur
Solution Approach 1:
The patent implements dynamic event role generation sequence selection through the IPGPF method, where the system automatically adapts the generation sequence based on document characteristics rather than using fixed manual sequences. This dynamic approach allows the model to adjust to different document types and event structures, resolving the contradiction between achieving accurate extraction and maintaining performance stability.
Solution Approach 2:
The system performs self-service by automatically determining the optimal event role generation sequence without manual intervention. The IPGPF method enables the model to self-adapt to different documents by learning from training data, eliminating the need for manual sequence selection and thereby stabilizing performance across diverse inputs.
2Productivity
If parallel generation methods are used, then event extraction speed is improved, but under-fitting occurs
Solution Approach 1:
The patent applies preliminary action through the pre-filling strategy in IPGPF, where certain event roles are pre-filled based on initial predictions before the main parallel generation process. This preliminary step provides better initial conditions for the parallel generation, allowing it to converge to more accurate solutions while maintaining speed, thus resolving the under-fitting problem.
Solution Approach 2:
The method ensures continuity of useful action by combining parallel generation with iterative refinement. Instead of single-pass parallel generation that causes under-fitting, the system continuously refines event role assignments across multiple iterations, maintaining both speed and accuracy by keeping the extraction process active and adaptive throughout.
3Adaptability or versatility
If event arguments are dispersed across multiple sentences, then comprehensive event extraction is possible, but extraction accuracy decreases
Solution Approach 1:
The patent introduces an intermediary mechanism through document-level context modeling in IPGPF. This intermediary layer maintains and propagates contextual information across sentence boundaries, allowing the system to connect dispersed event arguments while preserving accuracy. The context mediator bridges the gap between scattered arguments and coherent event reconstruction.
Solution Approach 2:
The method achieves universality by creating a document-level event extraction framework that handles both localized and dispersed arguments through the same IPGPF mechanism. This universal approach works regardless of whether arguments are concentrated in one sentence or distributed across multiple sentences, maintaining consistent accuracy across different document structures.
Data Source
AI summary
Embodiments of the present disclosure relate to a method, an apparatus, an electronic device and a medium for event extraction. The method comprises: extracting a plurality of named entities from a document as a plurality of event arguments. The method further includes determining an event type and a template corresponding to the event type in the document. The method also includes filling the plurality of event arguments in respective locations in the template to generate a plurality of candidate event records, and filtering the plurality of candidate event records to obtain one or more target event record. In this way, respective candidate event records are generated through iteration while event extraction is performed at a document level, which can avoid performance fluctuations caused by manual selection of an event role generation sequence, and can also avoid under-fitting brought about by parallel generation to thus improve the event extraction accuracy.