Reproduce “Explainable Sentiment Analysis with DeepSeek-R1: Performance, Efficiency, and Few-Shot Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Donghao Huang, Zhaoxia Wang

Researchers from Singapore Management University evaluated DeepSeek-R1, an open-source large language model, for explainable sentiment analysis, comparing its performance and few-shot learning capabilities against proprietary models like GPT-4o. DeepSeek-R1 achieved higher accuracy, for instance, an F1 score of 91.39% on five-level sentiment classification, and offered unique dual-layer explanations that trace its detailed decision-making process.

Key insights

  • DeepSeek-R1's native reasoning architecture provides superior few-shot learning efficiency, often reaching peak performance with significantly fewer examples than proprietary models.
  • The quality of DeepSeek-R1's distilled variants is more dependent on the base architectural design (e.g., Qwen2.5-based 32B variant outperformed Llama-based 70B) than on raw parameter count, challenging parameter-centric assumptions.
  • DeepSeek-R1's unique dual-layer explanation, comprising a concise explanation and an extensive step-by-step reasoning trace, offers enhanced transparency by revealing the full decision-making process, including self-correction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.