DuPa-ASA

Dual-Path Attention-Based Sentiment Analysis Model

NOTIFICATION:

Note : Model need interface.

Note: The README still need a further revise and implements.

Note: The test parameter need to be appointed.

TO-DO:

Combine transfer study. (BERT, XLNet)
Ablation Study. (Next week)
Need more comprehensive analysis of the model's sensitivity to different parameter settings.

DataSet

IMDB

Yelp

Amazon

python run.py --model=BLAT --embedding=random --word True

Without transfer learning，seed:88 (M40单卡):

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
BiLSTM	0.5098	0.7081	0.6316	0.5061	0.5889
TextCNN	0.8702	0.9681	0.7735	0.8947	0.87663
TextCNN_Att	0.8644	0.9639	0.7137	0.8756	0.8544
TextRCNN	0.7480	0.9267	0.7186	0.8377	0.8078
Transformer	0.7746	0.9344	0.7110	0.8093	0.8073
DPCNN	0.8742	0.9723	0.7796	0.9058	0.8830
Fasttext	0.8766	0.9534	0.7380	0.8517	0.8549
Fastformer	0.8718	0.9664	0.7739	0.8813	0.8734
BLAT	0.8630	0.9731	0.7861	0.9119	0.8835

With transfer learning(seed 88)(M40单卡):

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
bert(bz=64, S)	0.8721	0.9562	0.7495	0.8446（.4,DP)
xlnet(bz=64,S)	0.8955	0.9572	0.7541	0.8601	0.8667
BLAT(bert)	0.8954	0.9744	0.7915	0.9188	0.8950
BLAT(xlnet)	0.8874	0.9748	0.7892	0.9215	0.8932

seed:3407,bz=128, 2080ti，单卡

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
BLAT	0.8772	0.9729	0.7861	0.9127	0.8872
BLAT(bert)	0.9000	0.9749	0.7880	0.9212	0.8960
BLAT(xlnet)	0.8884	0.9751	0.7891	0.9238	0.8941
BLAT(bert-large)	0.8914	0.9759	0.7915	0.9226	0.8954
BLAT(xlnet-large)	0.8860	0.9761	0.7872	0.9209	0.8926

Ablation Study: seed=88, bz=128, 单卡(M40)

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
Fastformer(bert_base)	0.8734	0.9661	0.7732	0.8834	0.8740
TextCNN(bert_base)	0.9054	0.9758	0.7930	0.9213	0.8989

Ablation Study: seed=3407, bz=128, 单卡(M40)

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
Fastformer(bert_base)	0.8806	0.9664	0.7739	0.8827	0.8759
TextCNN(bert_base)	0.8628	0.9761	0.7930	0.9210	0.8882

seed:88, bz=64, 2080ti, 多卡

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
bert-large
xlnet-large

seed:88, bz=64, 2080ti, 多卡

Model	IMDB	Yelp-2	Yelp-5	Amazon	Average
bert-base+extract

Please cite our paper if you use it in your work:

@inproceedings{,
   title={{}: },
   author={},
   booktitle={},
   year={}
}