This academic article discusses a novel framework designed to enhance the natural language capabilities of Code Large Language Models (Code LLMs). Most existing works have focused on enhancing code LLMs from the programming capabilities perspective, with less attention paid to their natural language understanding abilities. The proposed framework, comprising two modules: AttentionExtractor and AttentionCoder, seeks to address this gap. The AttentionExtractor extracts key phrases from the user’s natural language requirements, while the AttentionCoder uses these phrases to generate target code. This framework integrates code LLMs with traditional natural language processing tools. The authors validate the framework using a new code generation benchmark, MultiNL-H, covering five natural languages. The experimental results demonstrate the effectiveness of the proposed framework.
Publication date: 26 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.14242
