Regex Embeddings for Compact Named Entity Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing named-entity recognition (NER) systems based on dictionaries require large storage space and are limited in recognizing patterns and similarities in structured or unstructured datasets, necessitating a more efficient and succinct method for data type classification.
Innovation Solution
Utilizing regular expression (regex) patterns to identify and generate vectors that represent the occurrence of specific patterns in datasets, which are then normalized and used to improve the accuracy of NER systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dictionary-based word embedding is used, then text can be converted into numerical representations, but storage space requirements increase significantly
Solution Approach 1:
The patent extracts only the essential pattern-matching information from full dictionary-based embeddings by using regular expressions to identify and represent recurring text patterns. This extraction approach captures the necessary structural information for NER while discarding redundant data, thereby reducing storage requirements while maintaining representation effectiveness.
Solution Approach 2:
Instead of converting words to vectors using comprehensive dictionaries, the patent inverts the approach by using regular expressions to directly identify patterns in text and represent them numerically. This inversion bypasses the need for large pre-computed embedding dictionaries while achieving comparable or superior pattern recognition for NER tasks.
2Adaptability or versatility
If comprehensive dictionaries are used for word embedding, then more words can be represented, but the system becomes less efficient and more complex
Solution Approach 1:
The patent makes the regular expression patterns universal by designing them to match recurring structural patterns across diverse text types and domains. These patterns can represent multiple different words and entities through a unified framework, providing broad adaptability without requiring domain-specific dictionaries for each application.
Solution Approach 2:
The patent replaces the mechanical dictionary-lookup system with a pattern-matching system based on regular expressions. This substitution eliminates the need for large static dictionaries and their associated complexity, while providing a more flexible and computationally efficient approach to text representation that adapts dynamically to different text inputs.
3Reliability
If traditional NER systems are used, then named entities can be identified, but pattern recognition and similarity detection in structured or unstructured datasets are limited
Solution Approach 1:
The patent segments text into recurring patterns using regular expressions, breaking down complex text structures into identifiable, representable units. This segmentation enables the system to recognize and represent patterns at multiple levels of granularity, improving both entity identification and the ability to detect similarities across different text structures.
Solution Approach 2:
The patent changes the parameters of text representation by using regular expression match counts and pattern frequencies as numerical features instead of traditional word vectors. This parameter transformation enables more effective pattern recognition and similarity detection while maintaining reliable named entity identification through the structured pattern-based representation.
Data Source
AI summary
Various embodiments are directed to a system that utilizes regular expression (regex) to recognize at least portions of characters, words, text, numbers, etc. in a structured or unstructured dataset, any patterns associated therewith, and/or similarities between the determined patterns. In examples, a regex-based pattern recognition platform may receive a dataset and determine whether at least a first regex pattern and a second regex pattern can be identified. The occurrences of the first and second regex patterns and the frequency of those occurrences may reveal something about the dataset itself or any patterns contained therein.