yskn67 / reading_papers

reading papers memorandum
MIT License
1 stars 0 forks source link

Autoencoding Keyword Correlation Graph for Document Clustering #16

Open yskn67 opened 4 years ago

yskn67 commented 4 years ago

一言でいうと

文章のembeddingをgraph autoencoderを使用してlocal contextとglobal contextの両方を埋め込む手法を提案 contributionは文章のグラフ化手法Keyword Correlation Graph(KCG)の構築

スクリーンショット 2020-07-26 4 33 08

KCGは以下の4ステップで構成されている

  1. NMFを使用して50個の特徴語を抽出
  2. 文章に含まれる文ごとに特徴語を1つ付与し、特徴語をNodeとする(1Nodeに複数の文が所属する)
  3. 文ごとにembeddingを生成する
  4. Node間にEdgeを張る

3のembeddingではSentence-BERTが使用され、4のEdgeは文章集合間の類似度に基づき構築された こちらの手法がクラスタリングにおいて既存手法をoutperform

提案手法ではgraph autoencoderとしてmulti-task graph autoencoderが使用されていた こちらをautoencoder, variational graph autoencoderと入れ替えたところ前者はlocal context、後者はglobal contextを掴んでいるとその後の分析で述べている よって、それらをoutperformしたmulti-task graph autoencoderはlocalとglobalなcontext両方掴んでいると主張したい模様

論文リンク

https://www.aclweb.org/anthology/2020.acl-main.366/

著者/所属機関

Billy Chiu† Sunil Kumar Sahu† Derek Thomas† Neha Sengupta† Mohammady Mahdy†

†Inception Institute of Artificial Intelligence, Abu Dhabi, United Arab Emirates

投稿日付(yyyy/MM/dd)

2020