magicpieh28 / Paper-Summary

1 stars 0 forks source link

Recursive Neural Conditional Random Fields for Aspect-based Sentiment Analysis(2016) #15

Open magicpieh28 opened 5 years ago

magicpieh28 commented 5 years ago

about this paper

Author: Wenya Wang, Sinno Jialin Pan, Daniel Dahlmeier, Xiaokui Xiao Link: http://aclweb.org/anthology/D16-1059

文章ごとに依存ツリーを生成するRNNから出たアウトフットをCRFに入れてaspect-basedな感情を表す単語を探索する手法。

aim

problems

model

image

image

基本的にheavy feature(あらかじめ用意されたレキシコン情報・構文論的分析・品詞タグなど)はいらないが、light feature(品詞タグ(Stanford POS tagger)・名前のリスト(訓練データから抽出した名前のリストセット2つ)・レキシコン情報(Hu and Liu, 2004a))を足すことでより良い精度を出すことも可能。

*個人的意見:heavy featureとlight featureで言及された品詞タグとレキシコンは一体何が違うんだ?結局あの情報使った方が良い制度を出すのに役立つという話だよね?

experiment

datasets

SemEval Challenge 2014 task 4: laptop, restaurant

settings

ベクトル初期化のために単語embeddingをword2vec(Yelp Challenge dataset)で訓練 batch size: 25 optimizer: AdaGrad(0.02) epoch: 5 learning rate: 0.02 window size: 3(laptop), 5(restaurant)

results

image

image

aspect抽出タスクでは、レストランドメインではベクトル次元が325のとき最も性能が良く、ノートパソコンドメインでは300のとき最も性能が良かった。 opinion抽出タスクでは、レストランドメインの場合ベクトル次元が75以上であるとき、ノートパソコンドメインの場合125の時が最も性能が良かった。

code

なし

next

Unsupervised word and dependency path embeddings for aspect term extraction. In IJCAI. 2016

Kralifornia commented 4 years ago

Dear magicpieh28,

I am intrigued by the method described in your paper, and I would like to run your RNCRF on an own dataset, and subsequently use the aspect-based sentiments as input for a recommender system approach proposed by me. I will use the output of two aspect-based sentiment analysis methods, and would like to use yours as one of these two. I perform this study for my Master's thesis of the Erasmus University in Rotterdam and ofcourse I will provide a reference in my thesis.

Please let me know whether this is possible!

Kind regards, Paul