Graph-Based Information Extraction via Joint Neural Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current information extraction systems perform entity recognition, relation extraction, and coreference resolution as separate tasks, leading to cascading errors and a lack of integration, while joint systems fail to perform coreference resolution and rely on word-level information, neglecting interactions between sentences.
Innovation Solution
An integrated end-to-end system that generates word-level, entity-level, and sentence-level graphs using a bidirectional long short-term memory and graph convolution network to identify entity candidates, extract links, and perform deconvolution to model linkages between sentences, enabling automated graph-based information extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If pipeline-based information extraction systems are used to perform entity recognition, relation extraction, and coreference resolution separately, then each task can be implemented in dedicated modules, but cascading errors occur and knowledge from one module is unavailable to assist other tasks
Solution Approach 1:
The patent merges entity recognition, relation extraction, and coreference resolution into a unified neural network model that processes all three tasks simultaneously. The model shares underlying representations and learns joint patterns across tasks, eliminating the modular separation that causes cascading errors while maintaining the functional distinctions of each extraction task through shared hidden layers and coordinated output heads.
2Reliability
If joint information extraction systems are used to simultaneously identify entities and relations, then output of one module is available to assist other tasks, but coreference resolution is not performed and word-level information is relied upon rather than higher levels
Solution Approach 1:
The patent extends traditional word-level and sentence-level processing by introducing a discourse-level dimension that spans multiple sentences. This enables the model to perform coreference resolution across sentence boundaries by maintaining and processing contextual representations at the discourse level, thereby adding a new dimensional layer of information integration that captures long-range dependencies and anaphoric relationships.
3Reliability
If graph neural network based end-to-end systems are used for information extraction, then entity recognition and relation extraction can be performed integratedly, but interactions between sentences are rarely considered when performing relation extraction
Solution Approach 1:
The patent addresses the limitation of sentence-isolated processing by extending the graph neural network to operate at the discourse level, where nodes represent discourse entities spanning multiple sentences and edges capture relationships across sentence boundaries. This dimensional extension allows the model to consider interactions between sentences while maintaining the benefits of end-to-end integrated extraction through shared representations and joint optimization.
Data Source
AI summary
A set of sentences within a natural language text document are parsed, generating a word-level graph corresponding to a sentence in the set of sentences. Within the word-level graph using a trained entity identification model, a set of entity candidates are identified. From a set of graphs modelling relationships between portions of the set of sentences, a set of embeddings is generated. From a set of pairs of embeddings in the set of embeddings using a set of deconvolution layers, a set of links between entity candidates within the set of entity candidates is extracted. From the set of links and the set of entity candidates, an output graph modelling linkages between portions of the set of sentences within the natural language text document is generated.