Open yskn67 opened 4 years ago
topic modelのようなことをNNで行なった論文。 特にレビューなどの短い文章を対象としている。 モデルはword embeddingのattention重み付き和をsentence embeddingとし、 それをautoencoderのようにreconstractすることでaspect(topic)への重みを得る。
いくつかのデータセットでaspect(topic)の一貫性という側面でoutperform。 attentionの有効性も評価。
本筋とはずれるがtopic modelよりもword2vecをk-meanでクラスタリングした方が 一貫性という点では良好だったのが少し面白かった。
https://www.aclweb.org/anthology/P17-1036/
Ruidan He†‡ Wee Sun Lee† Hwee Tou Ng† Daniel Dahlmeier‡
†Department of Computer Science, National University of Singapore ‡SAP Innovation Center Singapore
2017
一言でいうと
topic modelのようなことをNNで行なった論文。 特にレビューなどの短い文章を対象としている。 モデルはword embeddingのattention重み付き和をsentence embeddingとし、 それをautoencoderのようにreconstractすることでaspect(topic)への重みを得る。
いくつかのデータセットでaspect(topic)の一貫性という側面でoutperform。 attentionの有効性も評価。
本筋とはずれるがtopic modelよりもword2vecをk-meanでクラスタリングした方が 一貫性という点では良好だったのが少し面白かった。
論文リンク
https://www.aclweb.org/anthology/P17-1036/
著者/所属機関
Ruidan He†‡ Wee Sun Lee† Hwee Tou Ng† Daniel Dahlmeier‡
†Department of Computer Science, National University of Singapore ‡SAP Innovation Center Singapore
投稿日付(yyyy/MM/dd)
2017
概要
新規性・差分
手法
結果
コメント