Open magicpieh28 opened 5 years ago
Author: Peng Chen, Zhongqian Sun, Lidong Bing, Wei Yang Link: http://aclweb.org/anthology/D17-1047
レビュー文の中で示される感情表現が指しているaspectを探索するタスク。 bi-LSTM(GRUを用いており、input memory修正済み)とattention(レイヤーの数を増やしながら実験)のモデルを提案
labor-intensive特徴エンジニアリングをする必要がある
訓練コーパスが十分用意されないとターゲット単語と遠くにある感情表現はLSTMでは取りにくくなり、結局は特徴を失ってしまう
特定ドメインの訓練コーパスはツイッターなどでのみ学習がうまくできてる
LSTMの諸問題
attentionの諸問題
SemEval 2014(Laptop, Restaurant) collection of tweets(Dong et al., 2014) Chinese news comments
あらかじめ訓練されたGloVeの300次元の単語embedding使用 中国語:CBOWと1.4billionのトークンを持っているコーパスで訓練 訓練iteration:100 attentionレイヤー数:1〜5 embedding:チューニングしたものとしなかったもの
RAM-3AL-T-R: あらかじめ訓練されたembeddingは使用しない、ランダムに初期化、訓練時にチューニング RAM-3AL-T: あらかじめ訓練されたembeddingを初期値として使用、訓練時にチューニング RAM-3AL-NT: あらかじめ訓練されたembeddingをしようしたが訓練時にチューニングは行なっていない
ない
Clause sentiment identification based on convolutional neural network with context embedding. In ICNC-FSKD. 2016 On the properties of neural machine translation: Encoder-decoder approaches. CoRR. 2014 Phrase RNN: Phrase recursive neural network for aspect-based sentiment analysis. In Proceedings of 2015 EMNLP
about this paper
Author: Peng Chen, Zhongqian Sun, Lidong Bing, Wei Yang Link: http://aclweb.org/anthology/D17-1047
レビュー文の中で示される感情表現が指しているaspectを探索するタスク。 bi-LSTM(GRUを用いており、input memory修正済み)とattention(レイヤーの数を増やしながら実験)のモデルを提案
aim
problems
labor-intensive特徴エンジニアリングをする必要がある
訓練コーパスが十分用意されないとターゲット単語と遠くにある感情表現はLSTMでは取りにくくなり、結局は特徴を失ってしまう
特定ドメインの訓練コーパスはツイッターなどでのみ学習がうまくできてる
LSTMの諸問題
attentionの諸問題
how to solve
model
experiments
datasets
SemEval 2014(Laptop, Restaurant) collection of tweets(Dong et al., 2014) Chinese news comments
settings
あらかじめ訓練されたGloVeの300次元の単語embedding使用 中国語:CBOWと1.4billionのトークンを持っているコーパスで訓練 訓練iteration:100 attentionレイヤー数:1〜5 embedding:チューニングしたものとしなかったもの
results
RAM-3AL-T-R: あらかじめ訓練されたembeddingは使用しない、ランダムに初期化、訓練時にチューニング RAM-3AL-T: あらかじめ訓練されたembeddingを初期値として使用、訓練時にチューニング RAM-3AL-NT: あらかじめ訓練されたembeddingをしようしたが訓練時にチューニングは行なっていない
code
ない
next
Clause sentiment identification based on convolutional neural network with context embedding. In ICNC-FSKD. 2016 On the properties of neural machine translation: Encoder-decoder approaches. CoRR. 2014 Phrase RNN: Phrase recursive neural network for aspect-based sentiment analysis. In Proceedings of 2015 EMNLP