Weakly-Supervised Scientific Document Classification via Retrieval-Augmented Multi-Stage Training.

Area of Science:

• Computer Science
• Information Science
• Artificial Intelligence

Background:

• Scientific document classification is essential but hindered by high costs of human-labeled data.
• Existing methods struggle when label names contain domain-specific terms absent in the document corpus.

Purpose of the Study:

• To develop an effective scientific document classification method using only label names.
• To address the challenge of matching documents with semantically rich but potentially sparse label names.

Main Methods:

• Proposed WanDeR, a method employing dense retrieval for semantic matching in embedding space.
• Incorporated a label name expansion module to enrich label representations.
• Utilized a self-training step for refining classification predictions.

Main Results:

• WanDeR demonstrated superior performance compared to existing baselines.
• Achieved an 11.9% improvement in classification accuracy across three experimental datasets.
• The approach effectively captures label semantics for improved document matching.

Conclusions:

• WanDeR offers a cost-effective and accurate solution for scientific document classification.
• Leveraging dense retrieval and label expansion enhances the model's ability to understand label semantics.
• The proposed method shows significant potential for real-world applications in scientific information management.