Graph-Based Event Argument Extraction Network for Document-Level NLP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional event argument extraction systems are not scalable for document-level extraction, as they fail to effectively handle event trigger words and argument candidate words located far apart in different sentences or documents, lacking the incorporation of external knowledge and multi-hop heterogeneous reasoning.
Innovation Solution
An event argument extraction network is developed that generates rich document structures using multiple sources of information, including syntax, discourse, and external knowledge, employing a graph-based network to enable multi-hop and heterogeneous interactions, allowing for efficient document-level event argument extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sentence-level event argument extraction is used, then extraction accuracy for nearby words is improved, but scalability to document-level extraction deteriorates
Solution Approach 1:
The patent transitions from sentence-level to document-level extraction by adding the dimension of cross-sentence reasoning. The graph-based network enables multi-hop interactions that traverse sentence boundaries, allowing the system to capture relationships between event triggers and arguments located in different sentences or paragraphs throughout the entire document.
Solution Approach 2:
The patent segments the document into multiple sentences and represents each sentence as a separate node in the graph structure. This segmentation allows the system to process large documents in manageable units while maintaining the ability to reason across sentence boundaries through the graph connections, thus achieving both accuracy and scalability.
2Device complexity
If conventional extraction systems are used, then system simplicity is maintained, but ability to handle distant words deteriorates
Solution Approach 1:
The patent introduces an intermediary graph-based network structure that mediates between the input document and the extraction output. This graph structure serves as an intermediate representation that captures semantic relationships and enables multi-hop reasoning, improving the system's ability to handle distant word relationships while maintaining a relatively simple overall architecture.
Solution Approach 2:
The patent merges multiple information sources including syntax structures, discourse structures, and external knowledge bases into a unified graph-based representation. This combining of diverse structural information enhances the system's reliability in handling distant word relationships without requiring complex separate processing pipelines for each information type.
3Loss of information
If external knowledge and multi-hop reasoning are incorporated, then extraction completeness is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary action by pre-processing the document to extract syntax structures, discourse structures, and relevant external knowledge before the main extraction process. These pre-computed structures are stored and reused during the graph-based reasoning process, reducing redundant computations and managing computational complexity while ensuring extraction completeness.
Solution Approach 2:
The patent creates copied representations of the document in the form of graph structures that encode syntactic, discourse, and semantic relationships. Instead of repeatedly analyzing the original text during extraction, the system operates on these copied graph representations, which capture the essential structural information needed for complete argument extraction while reducing computational overhead.
Data Source
AI summary
Systems and methods for natural language processing are described. One or more embodiments of the present disclosure receive a document comprising a plurality of words organized into a plurality of sentences, the words comprising an event trigger word and an argument candidate word, generate word representation vectors for the words, generate a plurality of document structures including a semantic structure for the document based on the word representation vectors, a syntax structure representing dependency relationships between the words, and a discourse structure representing discourse information of the document based on the plurality of sentences, generate a relationship representation vector based on the document structures, and predict a relationship between the event trigger word and the argument candidate word based on the relationship representation vector.