Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Retrofitting Contextualized Word Embeddings with Paraphrases

katsutan
November 11, 2019

Retrofitting Contextualized Word Embeddings with Paraphrases

文献紹介
https://www.aclweb.org/anthology/D19-1113.pdf

長岡技術科学大学
勝田 哲弘

katsutan

November 11, 2019
Tweet

More Decks by katsutan

Other Decks in Technology

Transcript

  1. 文献紹介 RETROFITTING CONTEXTUALIZED WORD EMBEDDINGS WITH PARAPHRASES Proceedings of the

    2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1198–1203, Hong Kong, China, November 3–7, 2019. 長岡技術科学大学 勝田 哲弘
  2. EXPERIMENT  PARをELMoに適用し、広範な文レベルのタスクとadversarial SQuADコーパス で、改造されたELMoの品質を評価  パラメータの設定は論文を参照  Evaluation 

    SentEval framework  two baselines models:  (1) ELMo (all layers) 3,074-dimensional sentence embedding by averaging all layers.  (2) ELMo (top layers) 1,024 dimensional vector by averaging of the top layer. 8
  3. TASK DESCRIPTIONS  Sentence classification tasks  4つの文分類タスク(感情分析、製品レビュー、意見の極性)  Sentence

    inference tasks  言い換えの特定、テキストの含意  Semantic textual similarity tasks, Semantic relatedness tasks  0から5までの人間のラベル付けされたスコアに基づいて2つの文の意味的関連性の程度を測定  Adversarial SQuAD 9