Coreference Resolution via Selective Feature Activation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional coreference resolution processes are computationally demanding and time-consuming, especially when dealing with large-scale data, and struggle to maintain precision in cross-document resolution.
Innovation Solution
A computer-implemented method that assigns high-probability coreferent chains to high-confidence buckets to produce a Zipfian-like distribution, grouping chains based on context-based names and categories, and performing iterative global coreference resolution functions in a distributed computing environment to improve resolution efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional coreference resolution processes are applied to large-scale data, then resolution can be performed, but computational demand and time consumption increase excessively
Solution Approach 1:
The patent segments the coreference resolution process into multiple phases: mention extraction, chain formation, grouping into buckets based on context-based names and categories, and iterative resolution. This segmentation allows the system to handle large-scale data by processing it in manageable chunks rather than attempting to resolve all mentions simultaneously, thereby reducing time consumption while maintaining productivity.
Solution Approach 2:
The patent performs preliminary actions by extracting mentions and forming chains before performing the actual coreference resolution. It also pre-groups chains into buckets based on context-based names and categories, which prepares the data structure for more efficient processing. This preliminary organization reduces the computational complexity of subsequent resolution operations on large datasets.
2Productivity
If conventional coreference resolution processes are applied to large-scale data, then resolution can be performed, but computational demand increases excessively
Solution Approach 1:
The patent divides the computational task into segmented operations: mention extraction, chain formation, bucket grouping, and iterative resolution. By segmenting the computation, the system can process large-scale data with reduced computational demand at each stage, using distributed computing to further manage resource requirements.
Solution Approach 2:
The patent changes parameters by creating a Zipfian-like distribution of buckets based on frequency, which allows the system to focus computational resources on the most important entities. It also uses iterative resolution with convergence criteria that adjust computational intensity based on data characteristics, reducing overall computational demand while maintaining resolution accuracy.
3Measurement precision
If conventional coreference resolution processes are used, then resolution can be performed, but precision in cross-document resolution is difficult to maintain
Solution Approach 1:
The patent segments the resolution process into local coreference resolution within documents and global coreference resolution across documents. This segmentation allows the system to maintain precision in cross-document resolution by building on accurate local resolutions rather than attempting to resolve everything simultaneously, reducing the complexity burden at each stage.
Solution Approach 2:
The patent implements iterative resolution with feedback mechanisms where the system refines its resolutions based on convergence criteria and similarity comparisons. This feedback loop allows the system to improve precision in cross-document resolution by continuously adjusting and verifying its decisions, while the iterative nature provides a structured approach to managing processing complexity.
4Reliability
If all coreferent chains are processed uniformly, then complete resolution is achieved, but processing efficiency decreases
Solution Approach 1:
The patent applies local quality by creating different processing strategies for different parts of the data. It groups chains into buckets based on context-based names and categories, then applies iterative resolution with convergence criteria that adapt to each bucket's characteristics. This allows the system to process common entities efficiently while maintaining completeness for rare entities, balancing productivity and reliability.
Solution Approach 2:
The patent changes parameters by using frequency-based bucket creation that produces a Zipfian-like distribution, allowing the system to process high-frequency entities more efficiently while maintaining attention to low-frequency entities. The iterative resolution process adjusts computational intensity based on convergence progress, maintaining completeness while improving overall processing efficiency.
Data Source
AI summary
In some aspects, systems, methods, and computer-readable media for selective feature activation for coreference resolution are disclosed. In one embodiment, a method includes receiving text data comprising a plurality of mentions corresponding to entities, and determining a plurality of data features, comprising semantic features and syntactic features, for comparing a particular pair of mentions from the plurality of mentions. The method also includes selectively activating a subset of features from the plurality of data features based on semantic and syntactic context of the particular pair of mentions within the text data, and determining, using weights associated with the activated subset of features and at least one machine learning function, whether a first mention of the pair of mentions and second mention of the pair of mentions refer to a same entity.