Open yskn67 opened 4 years ago
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
KCGは以下の4ステップで構成されている
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
https://www.aclweb.org/anthology/2020.acl-main.366/
Billy Chiu† Sunil Kumar Sahu† Derek Thomas† Neha Sengupta† Mohammady Mahdy†
†Inception Institute of Artificial Intelligence, Abu Dhabi, United Arab Emirates
2020
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
KCGは以下の4ステップで構成されている
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