Reproduce “A Semantic Parsing Framework for End-to-End Time Normalization” on OpenResearch

About this paper

Xin Su, Sungduk Yu, Phillip Howard, Steven Bethard

An end-to-end framework for time normalization leverages an executable Python implementation of the SCATE framework and large language models for code generation. This approach enables smaller, deployable models to achieve competitive performance by using LLM-augmented data and notably outperforms direct interval generation.

Key insights

• Framing time normalization as SCATE code generation significantly improves performance compared to direct interval generation, leveraging LLMs' coding capabilities for better semantic understanding and deterministic output.
• An LLM-based data augmentation strategy with execution-based validation effectively overcomes data scarcity for complex semantic parsing, providing high-quality training examples at scale.
• Small, locally deployable language models, when fine-tuned on a combination of LLM-augmented and human-annotated data, can achieve performance competitive with or surpass larger generative LLMs.

Summary by alphaXiv.