Multi-label Sequential Sentence Classification via Large Language Model.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Bioinformatics

Background:

• Sequential Sentence Classification (SSC) is vital for scientific information retrieval and summarization.
• Existing SSC methods face limitations in model size, sequence length, and single-label classification.
• There is a need for advanced SSC approaches that can handle complex, multi-label scientific text data.

Purpose of the Study:

• To propose LLM-SSC, a large language model (LLM)-based framework for both single- and multi-label SSC tasks.
• To enhance task understanding in SSC by leveraging LLMs with designed prompts, demonstrations, and prediction target queries.
• To introduce a novel multi-label contrastive learning loss with an auto-weighting scheme for improved multi-label classification.

Main Methods:

• Utilized large language models (LLMs) for generating SSC labels via designed prompts.
• Incorporated demonstrations and a query within prompts to improve LLM task understanding.
• Developed and applied a multi-label contrastive learning loss with an auto-weighting scheme.
• Introduced the BIORC800 dataset for multi-label SSC analysis in the biomedical domain.

Main Results:

• LLM-SSC demonstrated strong performance in sequential sentence classification tasks.
• The framework showed effectiveness in both in-context learning and task-specific tuning settings.
• The proposed multi-label contrastive learning loss improved multi-label classification capabilities.

Conclusions:

• LLM-SSC offers a powerful and flexible framework for single- and multi-label SSC in scientific publications.
• The use of LLMs with prompt engineering significantly enhances SSC performance.
• The new BIORC800 dataset and contrastive learning approach advance multi-label SSC research, particularly in the biomedical domain.