CS224N lecture15 笔记

发布于 14 天前  31 次阅读

Natural language Generation


  • Decoding algorithms

    • Greedy decoding
    • Beam Search
    • Sampling methods
    • Softmax temperature
  • Summarization

    • extractive
    • abstractive
    • Summarization evaluation: ROUGE(Recall-Oridented Understudy for Gisting Evaluation)
      based on recall and has no brevity penalty(often a F1 version of ROUGE is reported)

      • ROUGE-1; ROUGE-2; ROUGE-L (Python implementation Slide 26)
    • copy mechanisms
    • Bottom-up summarization
      • Conten selection stage: Use a neural sequence-tagging model to tag words as include or don't include
      • Bottom-up attention stage: The seq2seq + attention system cannot attend to words tagged don't-include(apply a mask)
  • Dialogue

    • Negotiation dialogue
  • Conversational question answering: CoQA

  • Storytelling

    • EventToEvent story Generation
    • Skip-thought vectors
  • NLG evaluation

    • Word overlap based metrics(BLEU, ROUGE, METEOR, F1, etc.)