| Summary: | Generating high-quality word and sentence representations is a foundational task in natural language processing (NLP). In recent years, various embedding methodologies have been proposed, notably those leveraging the capabilities of large language models for in-context learning. Research has shown that language model performance can be enhanced by integrating a query with multiple examples. Inspired by this research, this project explores the use of a contrastive learning framework combined with ranking knowledge to enhance the generation and retrieval of sentence embeddings, aiming to more accurately identify the most similar sentences in in-context learning scenarios. Subsequent experiments tested various ranking strategies within the contrastive learning framework, yielding novel insights and conclusions.
|
|---|