Multi-Task Learning Framework for Text Classification Data Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning techniques face challenges in large-scale text classification tasks due to a lack of learning data and imbalance in data distribution across categories, leading to poor performance in sharing data effectively.
Innovation Solution
A method utilizing a multi-task learning framework based on an artificial neural network, where a shared layer is used to learn and classify text data across different scales, leveraging the Open Directory Project (ODP) for category data and applying a convolution neural network to enhance data sharing and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning technique is used for large-scale text classification, then classification performance is expected to improve, but performance deteriorates due to lack of learning data and data imbalance
Solution Approach 1:
The patent combines multiple classification tasks (small-scale and large-scale text classification) into a unified multi-task learning framework. By merging these tasks and sharing feature representations through common neural network layers, the system enables effective data sharing across tasks, allowing the large-scale classification to benefit from the abundant data in small-scale tasks while maintaining its own classification objectives
Solution Approach 2:
The patent creates a universal feature extraction framework that serves multiple classification tasks simultaneously. The shared neural network layers learn general text features that are applicable across different classification scales and domains, making the system multi-functional and enabling effective utilization of data from various sources for different classification objectives
2Quantity of substance
If explicit expression model is used for large-scale classification, then data sharing is improved through tree structure, but performance becomes dependent on knowledge base and cannot calculate semantic similarity
Solution Approach 1:
The patent replaces the mechanical explicit expression model (tree structure with predefined categories) with a neural network-based implicit representation system. This substitution enables the system to automatically learn semantic representations and calculate semantic similarity between texts through vector operations, while maintaining the data sharing capabilities through the hierarchical category structure integrated into the neural network
3Measurement precision
If deep learning-based multi-task learning is used for small-scale classification, then performance improvement is achieved, but large-scale classification cannot share data effectively
Solution Approach 1:
The patent extends the multi-task learning framework to an additional dimension by integrating hierarchical category structures into the neural network architecture. This dimensional extension allows the system to organize classification tasks at multiple levels (fine-grained and coarse-grained categories) and share data effectively across both small-scale and large-scale classification objectives through the hierarchical feature representations
Data Source
AI summary
According to an aspect of an embodiment, a method for a text classification using multi-task learning executed by a server, the method includes: a step of generating text classification learning data by executing a pre-processing process on raw data collected through a network; a step of learning a category classification in a first classification system preset through a first text classification model using the text classification learning data as an input; a step of learning a category classification in a second classification system preset through a second text classification model using the text classification learning data as an input; and a step of classifying the category in the second classification system by inputting text data received from a user terminal into the second text classification model.