Refinement Network for Granular Offensive Text Span Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional natural language processing systems fail to differentiate between individual words in a phrase and cannot accurately identify the offensive span within toxic text, relying on the existence of a target opinion and lacking the ability to pinpoint the location of offensive content.
Innovation Solution
A machine learning-based system that generates refined vector representations for words using a refinement network trained for offensive text recognition, allowing for the identification of offensive spans by generating label information that indicates whether each word is offensive, thereby pinpointing the location of toxic text without relying on a target opinion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional NLP systems are used for toxicity detection, then the system can predict sentiment polarity, but it cannot differentiate between individual words in a phrase or identify the offensive span location
Solution Approach 1:
The patent segments the text into individual words or tokens and processes each one separately through the neural network. Each word is represented as a vector and processed independently, allowing the system to identify which specific words are offensive rather than treating the entire phrase as a unit. This segmentation enables precise location identification of offensive spans.
Solution Approach 2:
The patent transforms the text from a one-dimensional sequence of words into a multi-dimensional vector space where each word is represented by a vector with multiple dimensions. This dimensional transformation allows the neural network to capture nuanced relationships and identify offensive content more precisely by analyzing the vector representations of individual words.
2Reliability
If the system relies on target opinion extraction to identify offensive content, then it can perform sentiment analysis, but it fails when no target opinion exists or cannot pinpoint the offensive span
Solution Approach 1:
The system performs self-service by directly analyzing each word's vector representation to determine offensiveness without requiring an external target opinion or reference. The neural network independently evaluates each word and generates labels indicating whether it is offensive, eliminating the dependency on target opinion extraction and improving reliability in edge cases.
Solution Approach 2:
The patent incorporates feedback mechanisms through the neural network's iterative processing, where the model continuously refines its predictions based on the vector representations and contextual information. This feedback loop enables the system to improve its accuracy in identifying offensive content without relying on pre-defined target opinions.
3Measurement precision
If the system processes text at the sentence level, then it can perform overall sentiment analysis, but it cannot identify specific offensive words or spans
Solution Approach 1:
The patent segments the text processing into individual word-level operations, where each word is processed separately through the neural network. This segmentation allows the system to identify offensive words precisely while maintaining efficient processing by handling each word independently rather than processing the entire sentence as a single unit.
Data Source
AI summary
Systems and methods for natural language processing are described. One or more embodiments of the present disclosure receive a span of text comprising an offensive span and a non-offensive span, generate a contextualized word embedding for each of a plurality of words of the span of text, generate a refined vector representation for each of the plurality of words based on the corresponding contextualized word embedding using a refinement network trained for offensive text recognition, generate label information for each of the plurality of words based on the corresponding refined vector representation, wherein the label information indicates whether each of the plurality of words includes offensive text, and transmit an indication of a location of the offensive span based on the label information.