Neural Network Hate Target Identification via Multi-Stage CNN-LSTM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for identifying hate targets in hate speech classification are ineffective due to high overlap of words among classes, lack of grammatical structure in tweets, class imbalance, and limited contextual information, making it difficult for traditional classifiers to produce accurate results.
Innovation Solution
A neural network system comprising a convolutional neural network (CNN) with long short-term memory (LSTM) architecture, a second CNN for processing secondary data sets, and a deep neural network (DNN) to combine and categorize text messages into relevant targets, utilizing graph-based and semantic features to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional classifiers are used for hate speech classification, then the classification process is simple, but the classification accuracy is low due to high word overlap among different hate target classes
Solution Approach 1:
The patent segments the classification task into multiple stages: first identifying hate targets (entities being targeted), then classifying the type of hate speech. This is achieved through separate neural network components - one for target identification and another for hate speech type classification, allowing each component to specialize and improve overall accuracy despite word overlap
Solution Approach 2:
The patent introduces additional dimensions for analysis by extracting multiple features including graph-based features (word co-occurrence networks), semantic features (word embeddings), and syntactic features. This multi-dimensional feature space allows the classifier to distinguish between overlapping hate target classes by considering relationships beyond individual word presence
2Measurement precision
If syntactic parsers and subject-verb-object analyzers are used to improve structural analysis, then grammatical features are enhanced, but the results are inaccurate due to tweets not following proper grammatical rules
Solution Approach 1:
The patent changes the approach from relying on traditional grammatical parameters (syntax trees, SVO structures) to using distributional parameters derived from graph-based features and word embeddings. This allows the system to capture semantic relationships in informal tweet language without requiring proper grammatical structure
3Measurement precision
If more contextual and background information is added to tweets, then the classification task becomes easier, but the data processing complexity and time increase
Solution Approach 1:
The patent performs preliminary processing by pre-computing graph-based features (word co-occurrence statistics) and semantic features (word embeddings) that can be reused across multiple classification tasks. This preliminary action reduces the computational burden during actual classification while still providing rich contextual information
4Measurement precision
If graph-based features and semantic features are extracted and processed, then the classification accuracy is improved, but the data processing complexity increases
Solution Approach 1:
The patent creates universal feature representations (graph-based features and semantic embeddings) that serve multiple functions: they capture semantic meaning, represent contextual relationships, and can be fed into different classification models. This multi-functionality reduces overall system complexity by using a single feature extraction pipeline for multiple purposes
Data Source
AI summary
Neural network systems are provided that comprise one or more neural networks. The first neural network can comprise a convolutional neural network (CNN) long short-term memory (LSTM) architecture for receiving a primary data set comprising text messages and output a primary data structure comprising a text pattern-based feature. The second neural network can comprise a CNN architecture for receiving a secondary data sets derived from the primary data set and output a plurality of secondary data structures. The third neural network can combine the data structures to produce a combined data structure, and then process it to produce a categorized data structure comprising the text messages assigned to targets. The primary data set can comprise hate speech and the categorized data structure can comprise target categories, for example, hate targets. Methods of operating neural network systems and computer program products for performing such methods are also provided.