Toward Robust Weakly Supervised Text Classification: Weak Supervision Generation and Correlation-Aware Supervision...

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Mining

Background:

• Multilabel text classification (MLTC) traditionally requires extensive labeled data, incurring high manual labor costs.
• Weakly supervised (WS) learning offers an alternative by utilizing predefined category words for supervision.
• Existing WS methods often struggle with noisy pseudolabels derived from category words.

Purpose of the Study:

• To propose a novel WS learning MLTC framework that reduces reliance on large labeled datasets.
• To develop a robust category word selection method to mitigate noise in weak supervision.
• To enhance pseudolabel accuracy through correlation-aware label propagation.

Main Methods:

• Category word selection with significance ranking and crowd-sourcing (Cws-src): A novel method for generating confident category words using a TW-ITF weighting scheme and manual selection.
• Weakly supervised multilabel text classification with correlation-aware label propagation (Wmltc-clp): A generic WS learning method that estimates accurate pseudolabels by graph-based propagation.
• Extensive experiments on nine benchmark datasets, including sentiment analysis and general MLTC datasets, to evaluate the framework's effectiveness.

Main Results:

• The Cws-src method successfully generates more confident category words compared to baseline approaches.
• The Wmltc-clp method demonstrates significant performance improvements over existing WS learning baselines.
• Maximum performance gains reached 0.096 (Micro-F1), 0.081 (Macro-F1), 0.075 (AP), and 0.02 (RL) across benchmark datasets.

Conclusions:

• The proposed WS MLTC framework effectively addresses the challenge of limited labeled data.
• The Cws-src and Wmltc-clp components contribute to improved accuracy and robustness in MLTC.
• This framework offers a cost-effective and efficient alternative for real-world MLTC applications.