Scholarlink

Sarcasm detection, a challenge in sentiment analysis crucial for Ambient Intelligence (AmI), faces hurdles with cloud LLMs due to privacy, cost, and accessibility. This paper evaluates locally runnable LLMs for detecting author-intended sarcasm using the iSarcasm dataset. Our evaluation of Gemma, Phi, Mistral, and Qwen variants shows Qwen (qwen3:30B) achieves the highest F1-score of 0.507 for sarcasm. These LLMs demonstrate improved performance over traditional methods (average F1-score \(\approx 0.332\) ), though their performance currently trails human annotation (F1-score of 0.616). Despite this difference, our findings highlight their significant potential for on-premise, privacy-preserving sentiment analysis in AmI, paving the way for more responsive and context-aware human-computer interaction.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enabling Context-Aware Sarcasm Detection in Ambient Intelligence Through Local LLMs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Adelino Gala,
  • Manuel Rodrigues,
  • Francisco S. Marcondes,
  • Paulo Novais

摘要

Sarcasm detection, a challenge in sentiment analysis crucial for Ambient Intelligence (AmI), faces hurdles with cloud LLMs due to privacy, cost, and accessibility. This paper evaluates locally runnable LLMs for detecting author-intended sarcasm using the iSarcasm dataset. Our evaluation of Gemma, Phi, Mistral, and Qwen variants shows Qwen (qwen3:30B) achieves the highest F1-score of 0.507 for sarcasm. These LLMs demonstrate improved performance over traditional methods (average F1-score \(\approx 0.332\) ), though their performance currently trails human annotation (F1-score of 0.616). Despite this difference, our findings highlight their significant potential for on-premise, privacy-preserving sentiment analysis in AmI, paving the way for more responsive and context-aware human-computer interaction.