Learning Few-shot Sample-set Operations for Noisy Multi-label Aspect Category Detection

Shiman Zhao, Wei Chen, Tengjiao WangIJCAI 2023 · DOI

Abstract

Multi-label Aspect Category Detection (MACD) is a critical task in sentiment analysis that identifies multiple fine-grained categories within a single text. However, few-shot MACD faces significant challenges due to data scarcity and semantic noise from overlapping labels. We present the Few-shot Sample-set Operation (FSO) framework, which utilizes neural arithmetic—intersection, subtraction, and union—to treat text features as sets. This approach allows the model to logically isolate category-specific signals and subtract irrelevant noise. Experimental results on the FewAsp dataset demonstrate that FSO significantly outperforms state-of-the-art methods, particularly in low-shot scenarios, by creating more discriminative class prototypes.

Introduction

Multi-label Aspect Category Detection (MACD) requires systems to tag text with multiple labels simultaneously, such as identifying "Service," "Money," and "Internet" in a single restaurant review. In many real-world scenarios, labeled data is scarce, requiring models to learn from only a few examples (few-shot learning).

A primary obstacle in few-shot MACD is the "noise problem." Standard prototypical networks create class representations by averaging sample vectors. If a training sample for "Service" also mentions "Food," the resulting prototype becomes contaminated with food-related data. This overlap makes it difficult for models to distinguish between similar categories when data is sparse.

Examples of multi-label aspect categories and the noise problem in few-shot learning.

Methodology

To address the noise problem, the proposed framework introduces Few-shot Sample-set Operations (FSO). Instead of simple averaging, the model performs "neural arithmetic" to isolate relevant features:

  • Intersection ($M_{int}$): Extracts shared semantic features between samples of the same category to identify the core "essence" of the class.
  • Subtraction ($M_{sub}$): Identifies and removes shared content between different categories to isolate and discard irrelevant noise.
  • Union ($M_{uni}$): Ensures consistency by verifying that the extracted signal and noise can be recombined to reconstruct the original sample.

System Architecture

The system employs a BERT-based encoder to transform text into vectors, using category descriptions (e.g., the word "Service") as prior knowledge to guide feature extraction. The FSO module then processes these vectors to generate discriminative prototypes. To handle multi-label classification, the model uses a Multi-label Inference module with a modified "Circle Loss," which allows it to learn flexible classification thresholds dynamically.

The overall architecture of the FSO framework, including the Sample-set Operation Network and Feature Vector Extraction Module.

Experimental Results

The FSO method was evaluated on the FewAsp dataset, which includes 100 aspect categories split into training, validation, and testing sets. The model was tested on both multi-label and random sentence distributions.

Key Findings

  • Performance: FSO significantly outperformed existing state-of-the-art models like LPN and LDF. In 2-shot and 3-shot scenarios, FSO improved the Macro-F1 score by an average of 8.59%, with peak improvements reaching 11.01%.
  • Complexity Robustness: As the number of possible classes increased (e.g., 20-way tasks), FSO maintained higher accuracy than competitors, proving its ability to handle fuzzy boundaries between many categories.
  • Visual Evidence: T-SNE visualizations confirmed that FSO creates tighter, more distinct clusters in the feature space compared to baseline methods, which often produce overlapping or poorly defined clusters.

Conclusion

The Few-shot Sample-set Operation framework offers a robust solution for noisy multi-label classification by replacing simple vector averaging with logical set operations. By effectively filtering noise and leveraging category-specific anchors, FSO achieves superior performance in data-scarce environments. Future work may explore applying these set operations to general sentence embeddings and cross-modal tasks where noise reduction is equally critical.

Summary updated

BibTeX
@inproceedings{zhao2023learning,
  author = {Shiman Zhao and Wei Chen and Tengjiao Wang},
  title = {{Learning Few-shot Sample-set Operations for Noisy Multi-label Aspect Category Detection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/589},
}
Figure