A new benchmarking corpus called STRABLE has been released to address the understudied area of tabular machine learning that includes string entries. The corpus contains 108 real-world tables and was used to conduct a large-scale empirical study evaluating 445 different machine learning pipelines. The study found that for categorical-dominant tables, simple string embeddings combined with advanced tabular learners perform well computationally, while free-text-dominant tables benefit from large LLM encoders, though performance varies by LLM family and post-processing. AI
IMPACT Establishes a new benchmark for tabular machine learning with strings, potentially guiding future research and development in this area.
RANK_REASON Publication of a new research paper and dataset. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →