論文筆記-Co-GAT: A Co-Interactive Graph Attention Network for Joint Dialog Act Recognition and Sentimen

Co-GAT: A Co-Interactive Graph Attention Network for Joint Dialog Act Recognition and Sentiment Classification

動機

• where dialog act and sentiment can indicate the explicit and the implicit intentions separately. SC can detect the sentiments in utterances which can help to capture speakers’implicit intentions.
  
• 認為兩個資訊很重要，上下文資訊和互動資訊，之前的方法要麼考慮一個資訊，要麼是使用pipeline的形式，單獨建模。
  

Related Work

• 1、上圖a coling 2018：Multi-task dialog act and sentiment recognition on Mastodon
  • We manually annotate both dialogues and sentiments on this corpus, and train a multi-task hierarchical recurrent network – joint learning
  • can implicitly extract the shared mutual interaction information, but fail to effectively capture the contextual information

• 2、上圖b PR期刊，只考慮上下文的資訊 Integrated neural network model for identifying speech acts, predicators, and sentiments of dialogue utterances
  • explicitly leverage the previous act information to guide the current DA prediction
  • the model ignores the mutual interaction information

資料中的一個例子:

• 3、2020 AAAI DCR-Net
  • capture the contextual information, followed then by a relation layer to consider the mutual interaction information.
  • Pipeline way：two info model independently

Contributions

• first attempt to simultaneously incorporate contextual information and mutual interaction information
• propose a co-interactive graph attention network where a cross-tasks connection and cross-utterances connection

Method

• Speaker-Level Encoder , 使用GNN考慮同一個說話的人的資訊。邊，如果是同一個人，為1
• Stacked Co-Interactive Graph Layer ，2N個結點，2N*2N條邊。
  • Cross-utt info
  • Cross-tasks info
    

Experiments

問題

• 本質上來說，該做法還是依賴資料中的某些規律。
• 跨任務圖的計算還是一種全連線的方式，有邊就是1，否則就是0；怎麼更好建模一階鄰居？