Span-Level ML Model for Entity and Assertion Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models face challenges in effectively identifying entities and determining assertions, such as moral or quality assertions, within unstructured text, particularly in detecting linguistic indicators of imminent adverse actions.
Innovation Solution
A machine learning model that uses a span-level approach, combining named entity recognition (NER) and assertion identification, where entities and assertions are represented as vectors in a multi-dimensional space, allowing for joint recognition and assertion regression based on candidate spans of tokens, leveraging backpropagation for training and inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate models are used for entity identification and assertion determination, then model simplicity is maintained, but analysis precision and capability to detect adverse actions are insufficient
Solution Approach 1:
The patent combines entity identification and assertion determination into a single integrated machine learning model. The model simultaneously performs both tasks by processing text inputs through shared embedding layers and generating both entity labels and assertion predictions in one unified framework, thereby improving analysis precision while managing complexity through architectural efficiency
Solution Approach 2:
The machine learning model is designed with multi-functionality to handle both entity recognition and assertion determination within a single system. The model processes text inputs and generates multiple types of outputs (entity classifications and assertion predictions) using a unified architecture that leverages shared representations and common processing pathways for both functions
2Measurement precision
If coarse-grained text analysis is used, then processing speed is maintained, but detection capability for imminent adverse actions is insufficient
Solution Approach 1:
The patent applies segmentation by dividing the text into candidate spans at the entity level rather than processing the entire text uniformly. The model identifies specific spans of text that contain entities and focuses assertion determination on these segmented portions, enabling finer-grained analysis of potentially adverse content while reducing computational load compared to analyzing the complete text corpus
3Reliability
If joint recognition and assertion regression is implemented, then capability to predict imminent threats is enhanced, but computational complexity increases
Solution Approach 1:
The model performs preliminary action by generating entity embeddings and identifying candidate spans before conducting assertion regression. This sequential approach within the unified model allows the system to prepare structured representations of entities and their contextual spans first, then apply assertion determination based on these pre-processed features, thereby managing computational complexity through staged processing
Data Source
AI summary
A computing machine receives an input comprising unstructured text. The computing machine identifies, within the unstructured text, one or more entities using a named entity recognition (NER) engine in a trained machine learning model. The trained machine learning model embeds tokens from the text into a vector space and uses generated embeddings to identify one or more tokens as being associated with the one or more entities. The computing machine determines, using the trained machine learning model that identifies the one or more entities and based on the embedded tokens, an assertion applied, within the text, to at least one entity. The assertion is represented as a vector in a multi-dimensional space. Each dimension corresponds to a part of the assertion. The trained machine learning model is a span-level model that both identifies the one or more entities and determines the assertion based on candidate spans of tokens.