SRM-CSR: unsupervised aspect category detection based on semantic-aware relevance modeling and contextual sentence...

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised aspect category detection identifies implicit topics in text without labeled data.
• Current methods often generate suboptimal pseudo-labels, hindering model training and performance.
• Existing approaches struggle with aspect discriminability and rely on general pre-trained models.

Purpose of the Study:

• To propose a novel framework (SRM-CSR) for high-quality pseudo-label generation in unsupervised aspect category detection.
• To integrate lexical-level aspect-relevant information and sentence-level contextual representations.
• To improve the accuracy and effectiveness of aspect category detection models.

Main Methods:

• Extracting Aspect-Relevant Terms (ARTs) based on domain specificity and semantic stability.
• Employing an entropy-driven mechanism to weight aspect terms within sentences.
• Utilizing Sentence-BERT for contextual sentence encoding and similarity computation.
• Generating pseudo-labels independently at lexical and sentence levels, then consolidating consistent results.
• Training a neural classifier using a post-trained Domain Knowledge BERT.

Main Results:

• The proposed SRM-CSR framework effectively generates high-quality pseudo-labels.
• The integrated approach significantly improves unsupervised aspect category detection.
• Achieved an average improvement of 2.9 percentage points in macro-F1 over baseline methods on three real-world datasets.

Conclusions:

• The novel pseudo-labeling strategy in SRM-CSR is effective for unsupervised aspect category detection.
• Combining lexical and sentence-level information enhances aspect term identification and model performance.
• The framework offers a robust solution for improving models trained without annotated data.