Transformer Encoder-Decoder for Semi-Structured Entity Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in reliably extracting information from communications that use non-standard syntax and flexible formatting, making it difficult to perform bulk data analysis and named entity recognition, especially when entities are not guaranteed to be in a specific order or format.
Innovation Solution
A system utilizing an encoder-decoder transformer architecture processes individual characters from a stream of semi-structured communications, generating vectors to accurately extract entities regardless of their order or syntax, and includes a method for generating robust training data by varying entity positions and formats.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If regular expressions or string keyword search are used to extract information from communications with nonstandard syntax, then the system is simple to implement, but the reliability of information extraction deteriorates due to flexible formatting and arbitrary entity ordering
Solution Approach 1:
The patent replaces mechanical string-matching systems (regular expressions, keyword search) with a neural network-based information extraction system. The neural network processes communications with nonstandard syntax and flexible formatting to reliably extract entities and their values, achieving high accuracy without relying on rigid syntactic patterns.
Solution Approach 2:
The patent changes the approach from fixed syntactic parameters (regex patterns, keyword positions) to learned semantic representations. By training the neural network on examples of communications with varying syntax and formatting, the system adapts to different parameter configurations and extracts information reliably across diverse formats.
2Productivity
If rule-based systems are used for named entity recognition, then the system is easier to implement, but parsing time increases and error rates increase due to inability to handle flexible formatting
Solution Approach 1:
The patent replaces slow, error-prone rule-based parsing with a neural network that processes communications in parallel. The neural network simultaneously handles multiple formatting variations and extracts entities accurately, achieving both high parsing speed and low error rates by eliminating the need for complex rule evaluation.
3Adaptability or versatility
If the system extracts entities in a fixed order, then the output is simple and consistent, but the system cannot handle communications where entities appear in arbitrary orders
Solution Approach 1:
The neural network is trained to handle multiple entity ordering patterns universally. It processes communications where entities may appear in any sequence and extracts them correctly, making the system adaptable to arbitrary formatting without requiring separate processing logic for each possible order.
Solution Approach 2:
The system dynamically adjusts its processing based on the input communication's structure. The neural network analyzes the specific formatting and ordering of entities in each communication and adapts its extraction strategy accordingly, rather than relying on a fixed expected order.
Data Source
AI summary
A system for named entity extraction from a stream of semi-structured communications is disclosed. The system comprises a server connected to a network through which the stream of semi-structured communications is transmitted. The server is configured to receive at least one communication from the stream of semi-structured communications, divide the at least one communication into a set of individual characters, and for each individual character, generate a vector of values associated with the individual character. Once generated, the server passes each generated vector through an encoder-decoder transformer architecture; and based on output of the encoder-decoder transformer architecture, extracts a plurality of values associated with entities in the at least one communication.