Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Explainable Sentiment Analysis with DeepSeek-R1: Performance, Efficiency, and Few-Shot Learning

Created by
  • Haebom

Author

Donghao Huang, Zhaoxia Wang

Outline

This paper presents the first comprehensive study comparing the open-source inference model DeepSeek-R1 with OpenAI's GPT-4o and GPT-4o-mini. We evaluated the performance of the 671B model and its scaled-down counterparts with just a few training runs, and found that DeepSeek-R1 achieved an F1 score of 91.39% on five emotion classification tasks and an accuracy of 99.31% on two emotion classification tasks. This represents an eight-fold improvement over GPT-4o, demonstrating high efficiency with just a few training runs. Furthermore, we analyzed the distillation effect by architecture, demonstrating that the 32B Qwen2.5-based model outperformed the 70B Llama-based model by 6.69 percentage points. DeepSeek-R1 improves explainability by transparently tracing the inference process step-by-step, but suffers from reduced throughput (Limitations).

Takeaways, Limitations

•
Takeaways:
◦
DeepSeek-R1 presents an open-source alternative model that achieves high accuracy with a few training rounds, much more efficiently than GPT-4o.
◦
We provide Takeaways for model development by analyzing the distillation effect according to the architectural features of DeepSeek-R1.
◦
Due to its high explainability through step-by-step tracking, it can contribute to the development of interpretable AI models.
•
Limitations:
◦
The step-by-step tracking process to ensure transparency of the inference process reduces throughput.
👍
Made with Slashpage