Deep-Learning-for-Aspect-Level-Sentiment-Classification-Baselines

The public state-of-the-art methods for deep learning based ASC. This is part of my survey paper "Deep learning for Aspect-level Sentiment Classification: Survey, Vision and Challenges".

Note

Click here to download all the ASC datasets (including SemEval 2014, SemEval 2015, SemEval 2016, Twitter, Sentihood, MPQA, Michell and manually-annoted-Hotel).

Authors

• Jie Zhou (jzhou@ica.stc.sh.cn), Jimmy Huang (jhuang@yorku.ca), Qin Chen, Tingting Wang, Qinmin Vivian Hu, and Liang He

• If you find this repo useful, please consider citing (no obligation at all):

@article{zhou2019deep,
  title={Deep learning for aspect-level sentiment classification: Survey, vision, and challenges},
  author={Zhou, Jie and Huang, Jimmy Xiangji and Chen, Qin and Hu, Qinmin Vivian and Wang, Tingting and He, Liang},
  journal={IEEE Access},
  volume={7},
  pages={78454--78483},
  year={2019},
  publisher={IEEE}
}

Environmental Requirement

• Python 3.6
• Pytorch 0.4.0
• sklearn
• numpy

Introduction of this work

models/: The deep learning model for ASC

• ContextAvg: the average of the word embeddings is fed to a softmax layer for sentiment prediction, which was adopted as a baseline in [1].
• AEContextAvg: the concatenation of the average of the word embeddings and the average of the aspect vectors is fed to a softmax layer for sentiment prediction, which was adopted as a baseline in [1].
• LSTM: the last hidden vector obtained by LSTM [2] is used for sentence representation and sentiment prediction.
• GRU: the last hidden vector obtained by GRU [3] is used for sentence representation and sentiment prediction.
• BiLSTM: the concatenation of last hidden vectors obtained by BiLSTM is used for sentence representation and sentiment prediction.
• BiGRU: the concatenation of last hidden vectors obtained by BiGRU is used for sentence representation and sentiment prediction.
• TD-LSTM: a target-dependent LSTM model which modeled the preceding and following contexts surrounding the target for sentiment classification [4].
• TC-LSTM: this model extends TD-LSTM by incorporating an target con- nection component, which explicitly utilizes the connections between target word and each context word when composing the representation of a sentence. [4].
• AT-LSTM: it uses an LSTM to model the sentence and a basic attention mechanism is applied for sentence representation and sentiment prediction. [5].
• AT-GRU: it uses a GRU to model the sentence and a basic attention mechanism is applied for sentence representation and sentiment prediction. [5].
• AT-BiLSTM: it uses a BiLSTM to model the sentence and a basic attention mechanism is applied for sentence representation and sentiment prediction. [5].
• AT-BiGRU: it uses a BiGRU to model the sentence and a basic attention mechanism is applied for sentence representation and sentiment prediction. [5].
• ATAE-LSTM: the aspect representation is integrated into attention-based LSTM for sentence representation and sentiment prediction [5].
• ATAE-GRU: the aspect representation is integrated into attention-based GRU for sentence representation and sentiment prediction.
• ATAE-BiLSTM: the aspect representation is integrated into attention-based BiLSTM for sentence representation and sentiment prediction.
• ATAE-BiGRU: the aspect representation is integrated into attention-based BiGRU for sentence representation and sentiment prediction.
• IAN: the attentions in the context and aspect were learned interactively for context and aspect representation [6].
• LCRS: it contains three LSTMs, i.e., left-, center- and right- LSTM, respectively modeling the three parts of a review (left context, aspect and right context) [7].
• CNN: The sentence representation obtained by CNN [8] is used for ASC.
• GCAE: it has two separate convolutional layers on the top of the embedding layer, whose outputs are combined by gating units [9].
• MemNet: the content and position of the aspect is incorporated into a deep memory network [10].
• RAM: a multi-layer architecture where each layer contains an attention-based aggregation of word features and a GRU cell to learn the sentence representation [11].
• CABASC: two novel attention mechanisms, namely sentence-level content attention mechanism and context attention mechanism are introduced in a memory network to tackle the semantic-mismatch problem [12].

data/: Store the data

• data_orign: the original datasets, including SemEval2014-Task4, SemEval2015-Task12, SemEval2016-Task5, Twitter, Sentihood, Michell, MPQA.
• data_processed: the datasets after processing
• store: Store the embedding of words, like GloVe.
• tmp: store the temporary files.

data_processing/: Processing the data

• SemEval2014-Laptop: Processe the Laptop14 dataset.
• SemEval2014-Resturant: Process the Restaurants14 dataset.
• SemEval2015-Resturant: Process the Restaurants15 dataset.
• SemEval2016-Resturant: Process the Restaurants16 dataset.
• Twitter: Process the Twitter dataset.
• MPQA: Process the MPQA dataset.
• Michell-en: Process the Michell-en dataset.
• Sentihood: Process the Sentihood dataset.

layers/: Basic units of deep learning models

• Attention: Attention units, including ''Contact Attention", ''General Attention" and ''Dot-Product Attention".
• Dynamic_RNN: Basic RNN, LSTM and GRU models.
• SqueezeEmbedding: Squeeze the embeddings of words.

results/: Store the results

• log: Store the log of the models.
• ans: Store the answer of the models
• attention_weight: Store the weight of the attentions.
• model: Store the trained models.

Statics of the performance of the existing works for deep learning based ASC

Method	Restaurants14		Laptop14		Restaurants15		Restaurants16		Twitter
	Accuracy	Marco-F1	Accuracy	Marco-F1	Accuracy	Marco-F1	Accuracy	Marco-F1	Accuracy	Marco-F1
RecNN for ASC
AdaRNN	-	-	-	-	-	-	-	-	66.30	65.90
PhraseRNN	66.20	-	-	-	-	-	-	-	-	-
RNN for ASC
GRNN	-	-	-	-	-	-	-	-	-	-
TD-LSTM	-	-	-	-	-	-	-	-	70.80	69.00
TC-LSTM	-	-	-	-	-	-	-	-	71.50	69.50
AE-LSTM	76.60	-	68.90	-	-	-	-	-	-	-
H-LSTM	-	-	-	-	-	-	-	-	-	-
Attention-based RNN for ASC
ATAE-LSTM	77.20	-	68.70	-	-	-	-	-	-	-
AB-LSTM	-	-	-	-	-	-	-	-	72.60	72.20
BILSTM-ATT-G	-	-	-	-	-	-	-	-	73.60	72.10
IAN	78.60	-	72.10	-	-	-	-	-	-	-
AF-LSTM(CONV)	75.44	-	68.81	-	-	-	-	-	-	-
HEAT	-	-	-	-	-	-	-	-	-	-
Sentic LSTM+TA+SA	-	-	-	-	-	-	-	-	-	-
PRET+MULT	79.11	79.73	71.15	67.46	81.30	68.74	85.58	79.76	-	-
PBAN	81.16	-	74.12	-	-	-	-	-	-	-
LSTM+SynATT+TarRep	80.63	71.32	71.94	69.23	81.67	66.05	84.61	67.45	-	-
MGAN	81.25	71.94	75.39	72.47	-	-	-	-	72.54	70.81
Inter-Aspect Dependencies	79.00	-	72.50	-	-	-	-	-	-	-
AOA-LSTM	81.20	-	74.50	-	-	-	-	-	-	-
LCR-Rot	81.34	-	75.24	-	-	-	-	-	72.69	-
Word&Clause-Level ATT	-	-	-	-	80.90	68.50	-	-	-	-
CNN for ASC
GCAE	77.28	-	69.14	-	-	-	-	-	-	-
PF-CNN	79.20	-	70.06	-	-	-	-	-	-	-
Conv-Memnet	78.26	68.38	76.37	72.10	-	-	-	-	72.11	70.80
TNet	80.69	71.27	76.54	71.75	-	-	-	-	74.97	73.60
Memory Network for ASC
MemNet	80.95	-	72.21	-	-	-	-	-	-	-
DyMemNN	-	58.82	-	60.11	-	-	-	-	-	-
RAM	80.23	70.80	74.49	71.35	-	-	-	-	69.36	73.85
CEA	80.98	-	72.88	-	-	-	-	-	-	-
DAuM	82.32	71.45	74.45	70.16	-	-	-	-	72.14	60.24
IARM	80.00	-	73.8	-	-	-	-	-	-	-
TMNs	-	68.84	-	67.23	-	-	-	-	-	-
Cabasc	80.89	-	75.07	-	-	-	-	-	71.53	-

The results of our implemented models

The results of dataset Restaurants14

	Accuracy	Macro			Micro			Precision			Recall			F1
		Precision	Recall	F1	Precision	Recall	F1	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.
ContextAvg	73.48	62.92	58.44	59.58	73.48	73.48	73.48	56.48	51.79	80.49	55.61	29.59	90.11	56.04	37.66	85.03
AEContextAvg	75.27	66.30	61.47	63.10	75.27	75.27	75.27	62.09	55.47	81.36	57.65	36.22	90.52	59.79	43.83	85.70
LSTM	77.23	67.54	64.34	65.51	77.23	77.23	77.23	63.35	54.55	84.73	61.73	39.80	91.48	62.53	46.02	87.98
GRU	78.75	70.51	65.61	67.11	78.75	78.75	78.75	67.36	59.84	84.35	66.33	37.24	93.27	66.84	45.91	88.58
BiGRU	77.14	67.61	63.55	65.15	77.14	77.14	77.14	64.94	53.69	84.19	57.65	40.82	92.17	61.08	46.38	88.00
BiLSTM	78.30	69.11	66.01	67.12	78.30	78.30	78.30	65.13	56.64	85.55	64.80	41.33	91.90	64.96	47.79	88.61
TD-LSTM	78.66	70.84	67.56	68.98	78.66	78.66	78.66	72.88	54.55	85.09	65.82	45.92	90.93	69.17	49.86	87.92
TC-LSTM	77.41	69.06	65.18	66.72	77.41	77.41	77.41	67.78	55.70	83.69	62.24	42.35	90.93	64.89	48.12	87.16
AT-LSTM	78.04	70.84	61.52	63.37	78.04	78.04	78.04	70.06	61.25	81.23	63.27	25.00	96.29	66.49	35.51	88.12
AT-GRU	78.30	70.74	64.76	66.58	78.30	78.30	78.30	67.91	61.21	83.11	64.80	36.22	93.27	66.32	45.51	87.90
AT-BiGRU	77.77	69.51	64.74	66.18	77.77	77.77	77.77	65.13	59.84	83.56	64.80	37.24	92.17	64.96	45.91	87.66
AT-BiLSTM	78.84	72.84	63.67	65.66	78.84	78.84	78.84	68.45	67.82	82.27	65.31	30.10	95.60	66.84	41.70	88.44
ATAE-GRU	76.79	68.68	63.49	65.32	76.79	76.79	76.79	69.49	54.62	81.92	62.76	36.22	91.48	65.95	43.56	86.44
ATAE-LSTM	76.79	67.93	62.74	63.72	76.79	76.79	76.79	64.53	57.00	82.25	66.84	29.08	92.31	65.66	38.51	86.99
ATAE-BiGRU	76.34	65.95	63.26	63.82	76.34	76.34	76.34	63.77	50.41	83.67	67.35	31.63	90.80	65.51	38.87	87.09
ATAE-BiLSTM	75.98	67.01	61.71	63.43	75.98	75.98	75.98	66.29	53.28	81.46	60.20	33.16	91.76	63.10	40.88	86.30
IAN	76.70	68.29	63.69	65.12	76.70	76.70	76.70	64.25	58.06	82.57	63.27	36.73	91.07	63.75	45.00	86.61
LCRS	76.25	68.71	60.85	63.03	76.25	76.25	76.25	69.82	56.44	79.88	60.20	29.08	93.27	64.66	38.38	86.06
CNN	75.18	68.45	58.44	60.25	75.18	75.18	75.18	60.44	65.79	79.12	56.12	25.51	93.68	58.20	36.76	85.79
GCAE	77.41	68.58	64.80	65.06	77.41	77.41	77.41	64.86	57.43	83.44	73.47	29.59	91.35	68.90	39.06	87.21
MemNet	73.39	62.74	61.13	61.09	73.39	73.39	73.39	52.56	52.38	83.29	62.76	33.67	86.95	57.21	40.99	85.08
RAM	77.41	68.38	65.67	66.76	77.41	77.41	77.41	67.20	53.25	84.68	64.80	41.84	90.38	65.97	46.86	87.44
CABASC	77.68	69.01	67.18	68.02	77.68	77.68	77.68	65.59	55.68	85.75	62.24	50.00	89.29	63.87	52.69	87.48

The results of Laptop14

	Accuracy	Macro			Micro			Precision			Recall			F1
		Precision	Recall	F1	Precision	Recall	F1	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.
ContextAvg	66.93	63.47	59.98	58.19	66.93	66.93	66.93	46.41	67.65	76.35	65.62	27.22	87.10	54.37	38.82	81.37
AEContextAvg	66.46	61.64	59.56	58.04	66.46	66.46	66.46	47.40	61.54	75.97	64.06	28.40	86.22	54.49	38.87	80.77
LSTM	66.14	62.37	60.20	55.35	66.14	66.14	66.14	48.08	62.79	76.23	78.12	15.98	86.51	59.52	25.47	81.04
GRU	67.71	64.31	61.50	58.60	67.71	67.71	67.71	49.47	66.67	76.80	73.44	23.67	87.39	59.12	34.93	81.76
BiGRU	69.44	65.61	63.83	61.49	69.44	69.44	69.44	49.22	67.11	80.49	74.22	30.18	87.10	59.19	41.63	83.66
BiLSTM	68.81	63.41	63.56	62.09	68.81	68.81	68.81	50.28	59.05	80.90	69.53	36.69	84.46	58.36	45.26	82.64
TD-LSTM	68.50	62.66	62.98	61.87	68.50	68.50	68.50	47.70	57.63	82.66	64.84	40.24	83.87	54.97	47.39	83.26
TC-LSTM	67.08	62.02	62.66	61.11	67.08	67.08	67.08	46.52	57.76	81.79	67.97	39.64	80.35	55.24	47.02	81.07
AT-LSTM	69.44	64.23	65.02	63.16	69.44	69.44	69.44	51.91	58.88	81.90	74.22	37.28	83.58	61.09	45.65	82.73
AT-GRU	70.85	66.57	66.21	63.58	70.85	70.85	70.85	54.21	64.63	80.87	80.47	31.36	86.80	64.78	42.23	83.73
AT-BiGRU	69.28	64.44	64.36	63.28	69.28	69.28	69.28	48.86	62.61	81.84	67.19	42.60	83.28	56.58	50.70	82.56
AT-BiLSTM	71.94	66.36	66.80	66.42	71.94	71.94	71.94	55.48	59.06	84.55	63.28	52.07	85.04	59.12	55.35	84.80
ATAE-GRU	69.75	64.43	63.46	62.45	69.75	69.75	69.75	52.76	61.22	79.31	67.19	35.50	87.68	59.11	44.94	83.29
ATAE-LSTM	67.40	65.16	62.18	58.47	67.40	67.40	67.40	47.39	69.64	78.44	78.12	23.08	85.34	59.00	34.67	81.74
ATAE-BiGRU	70.38	67.00	66.20	64.12	70.38	70.38	70.38	49.25	68.82	82.95	76.56	37.87	84.16	59.94	48.85	83.55
ATAE-BiLSTM	70.53	66.84	65.99	63.43	70.53	70.53	70.53	50.75	67.47	82.30	78.91	33.14	85.92	61.77	44.44	84.07
IAN	68.50	64.11	62.69	60.90	68.50	68.50	68.50	51.12	63.41	77.78	71.09	30.77	86.22	59.48	41.43	81.78
LCRS	66.46	63.15	60.84	59.50	66.46	66.46	66.46	46.70	66.67	76.08	66.41	33.14	82.99	54.84	44.27	79.38
CNN	66.93	65.95	59.91	57.75	66.93	66.93	66.93	45.99	76.36	75.51	67.19	24.85	87.68	54.60	37.50	81.14
GCAE	65.83	60.95	60.34	59.20	65.83	65.83	65.83	43.72	60.00	79.14	62.50	37.28	81.23	51.45	45.99	80.17
MemNet	64.42	59.08	59.36	58.01	64.42	64.42	64.42	43.01	54.87	79.35	62.50	36.69	78.89	50.96	43.97	79.12
RAM	67.55	62.25	60.78	59.73	67.55	67.55	67.55	49.09	60.44	77.23	63.28	32.54	86.51	55.29	42.31	81.60
CABASC	70.06	66.14	63.05	62.94	70.06	70.06	70.06	50.98	69.79	77.63	60.94	39.64	88.56	55.52	50.57	82.74

The results of Restaurants15

	Accuracy	Macro			Micro			Precision			Recall			F1
		Precision	Recall	F1	Precision	Recall	F1	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.
ContextAvg	72.31	65.35	50.18	49.80	72.31	72.31	72.31	74.91	50.00	71.15	58.67	2.22	89.65	65.80	4.26	79.34
AEContextAvg	73.37	49.87	50.22	49.17	73.37	73.37	73.37	78.54	0.00	71.06	59.25	0.00	91.41	67.55	0.00	79.96
LSTM	77.99	51.77	54.60	53.14	77.99	77.99	77.99	75.49	0.00	79.84	78.32	0.00	85.46	76.88	0.00	82.55
GRU	76.80	51.87	53.01	51.96	76.80	76.80	76.80	80.97	0.00	74.64	67.63	0.00	91.41	73.70	0.00	82.18
BiGRU	77.28	51.48	53.70	52.44	77.28	77.28	77.28	76.99	0.00	77.46	72.54	0.00	88.55	74.70	0.00	82.63
BiLSTM	78.34	52.34	54.36	53.14	78.34	78.34	78.34	79.18	0.00	77.84	72.54	0.00	90.53	75.72	0.00	83.71
TD-LSTM	77.28	64.53	57.65	59.04	77.28	77.28	77.28	78.06	37.50	78.04	71.97	13.33	87.67	74.89	19.67	82.57
TC-LSTM	74.44	62.62	53.41	54.10	74.44	74.44	74.44	76.51	37.50	73.84	65.90	6.67	87.67	70.81	11.32	80.16
AT-LSTM	80.00	53.32	55.82	54.48	80.00	80.00	80.00	79.88	0.00	80.08	78.03	0.00	89.43	78.95	0.00	84.50
AT-GRU	79.41	52.87	55.48	54.11	79.41	79.41	79.41	78.78	0.00	79.84	78.32	0.00	88.11	78.55	0.00	83.77
AT-BiGRU	77.99	61.04	54.64	54.30	77.99	77.99	77.99	81.88	25.00	76.24	70.52	2.22	91.19	75.78	4.08	83.05
AT-BiLSTM	79.88	53.11	55.88	54.45	79.88	79.88	79.88	78.41	0.00	80.93	79.77	0.00	87.89	79.08	0.00	84.27
ATAE-GRU	78.58	85.80	55.40	54.88	78.58	78.58	78.58	79.33	100.00	78.06	75.43	2.22	88.55	77.33	4.35	82.97
ATAE-LSTM	79.53	53.15	55.34	54.09	79.53	79.53	79.53	80.62	0.00	78.85	75.72	0.00	90.31	78.09	0.00	84.19
ATAE-BiGRU	78.70	69.08	56.30	56.29	78.70	78.70	78.70	77.46	50.00	79.80	77.46	4.44	87.00	77.46	8.16	83.25
ATAE-BiLSTM	78.34	52.21	54.59	53.29	78.34	78.34	78.34	78.14	0.00	78.47	75.43	0.00	88.33	76.76	0.00	83.11
IAN	79.41	86.18	56.74	56.82	79.41	79.41	79.41	78.82	100.00	79.72	77.46	4.44	88.33	78.13	8.51	83.80
LCRS	75.50	59.03	53.63	53.73	75.50	75.50	75.50	76.28	25.00	75.81	68.79	4.44	87.67	72.34	7.55	81.31
CNN	69.35	64.71	47.47	46.93	69.35	69.35	69.35	77.46	50.00	66.67	47.69	2.22	92.51	59.03	4.26	77.49
GCAE	76.33	57.61	53.89	53.32	76.33	76.33	76.33	75.07	20.00	77.76	73.99	2.22	85.46	74.53	4.00	81.43
MemNet	76.45	71.93	56.34	57.97	76.45	76.45	76.45	76.90	62.50	76.39	70.23	11.11	87.67	73.41	18.87	81.64
RAM	76.21	51.23	52.71	51.62	76.21	76.21	76.21	79.00	0.00	74.68	68.50	0.00	89.65	73.37	0.00	81.48
CABASC	76.21	61.73	56.28	57.30	76.21	76.21	76.21	76.47	31.25	77.47	71.39	11.11	86.34	73.84	16.39	81.67

The results of Restaurants16

	Accuracy	Macro			Micro			Precision			Recall			F1
		Precision	Recall	F1	Precision	Recall	F1	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.
ContextAvg	80.56	49.61	52.56	51.04	80.56	80.56	80.56	61.82	0.00	87.01	66.67	0.00	91.00	64.15	0.00	88.96
AEContextAvg	80.79	49.87	52.99	51.37	80.79	80.79	80.79	62.33	0.00	87.26	68.14	0.00	90.83	65.11	0.00	89.01
LSTM	83.12	76.89	59.24	58.23	83.12	83.12	83.12	64.23	75.00	91.43	81.86	6.82	89.03	71.98	12.50	90.22
GRU	83.47	69.29	60.53	61.34	83.47	83.47	83.47	67.81	50.00	90.07	77.45	13.64	90.51	72.31	21.43	90.29
BiGRU	83.47	77.36	59.66	61.39	83.47	83.47	83.47	68.18	75.00	88.91	73.53	13.64	91.82	70.75	23.08	90.34
BiLSTM	82.54	52.03	53.70	52.81	82.54	82.54	82.54	69.70	0.00	86.38	67.65	0.00	93.45	68.66	0.00	89.78
TD-LSTM	84.17	52.67	57.08	54.70	84.17	84.17	84.17	67.22	0.00	90.78	79.41	0.00	91.82	72.81	0.00	91.29
TC-LSTM	82.07	55.80	54.73	54.06	82.07	82.07	82.07	66.82	12.50	88.07	70.10	2.27	91.82	68.42	3.85	89.90
AT-LSTM	82.77	51.85	55.44	53.56	82.77	82.77	82.77	67.11	0.00	88.43	75.00	0.00	91.33	70.83	0.00	89.86
AT-GRU	83.82	52.68	56.04	54.30	83.82	83.82	83.82	69.06	0.00	88.99	75.49	0.00	92.64	72.13	0.00	90.78
AT-BiGRU	83.47	77.57	57.55	58.06	83.47	83.47	83.47	69.44	75.00	88.26	73.53	6.82	92.31	71.43	12.50	90.24
AT-BiLSTM	82.89	85.01	58.75	56.88	82.89	82.89	82.89	63.20	100.00	91.84	83.33	4.55	88.38	71.88	8.70	90.08
ATAE-GRU	82.31	51.16	55.11	53.01	82.31	82.31	82.31	64.41	0.00	89.09	74.51	0.00	90.83	69.09	0.00	89.95
ATAE-LSTM	82.19	51.70	53.10	52.33	82.19	82.19	82.19	69.07	0.00	86.02	65.69	0.00	93.62	67.34	0.00	89.66
ATAE-BiGRU	82.54	84.85	57.17	56.33	82.54	82.54	82.54	65.42	100.00	89.14	76.96	4.55	90.02	70.72	8.70	89.58
ATAE-BiLSTM	83.35	78.88	58.85	59.36	83.35	83.35	83.35	67.24	80.00	89.39	76.47	9.09	91.00	71.56	16.33	90.19
IAN	82.19	73.57	57.66	56.30	82.19	82.19	82.19	64.17	66.67	89.87	79.90	4.55	88.54	71.18	8.51	89.20
LCRS	81.61	68.60	57.16	59.36	81.61	81.61	81.61	70.31	50.00	85.50	66.18	13.64	91.65	68.18	21.43	88.47
CNN	81.84	73.19	55.21	55.14	81.84	81.84	81.84	65.44	66.67	87.48	69.61	4.55	91.49	67.46	8.51	89.44
GCAE	79.98	49.95	50.00	49.74	79.98	79.98	79.98	66.47	0.00	83.38	56.37	0.00	93.62	61.01	0.00	88.20
MemNet	81.26	67.07	55.25	57.94	81.26	81.26	81.26	70.41	46.15	84.64	58.33	13.64	93.78	63.81	21.05	88.98
RAM	83.47	52.61	55.12	53.83	83.47	83.47	83.47	70.00	0.00	87.83	72.06	0.00	93.29	71.01	0.00	90.48
CABASC	83.12	52.33	54.63	53.44	83.12	83.12	83.12	69.57	0.00	87.42	70.59	0.00	93.29	70.07	0.00	90.26

The results of Twitter

	Accuracy	Macro			Micro			Precision			Recall			F1
		Precision	Recall	F1	Precision	Recall	F1	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.	Neg.	Neu.	Pos.
ContextAvg	68.35	68.69	64.26	65.82	68.35	68.35	68.35	70.15	67.88	68.03	54.34	80.64	57.80	61.24	73.71	62.50
AEContextAvg	69.94	69.57	66.57	67.75	69.94	69.94	69.94	67.11	70.66	70.95	58.96	80.06	60.69	62.77	75.07	65.42
LSTM	69.22	69.64	65.13	66.52	69.22	69.22	69.22	66.46	69.12	73.33	63.01	81.50	50.87	64.69	74.80	60.07
GRU	68.79	67.37	68.11	67.71	68.79	68.79	68.79	64.32	73.80	64.00	68.79	70.81	64.74	66.48	72.27	64.37
BiGRU	67.20	67.63	62.14	63.68	67.20	67.20	67.20	67.11	66.74	69.03	58.96	82.37	45.09	62.77	73.74	54.55
BiLSTM	68.21	67.75	64.84	65.98	68.21	68.21	68.21	69.18	69.13	64.94	58.38	78.32	57.80	63.32	73.44	61.16
TD-LSTM	71.82	72.21	68.11	68.67	71.82	71.82	71.82	65.15	73.59	77.88	74.57	82.95	46.82	69.54	77.99	58.48
TC-LSTM	72.69	72.76	69.65	70.90	72.69	72.69	72.69	74.00	72.56	71.71	64.16	81.79	63.01	68.73	76.90	67.08
AT-LSTM	70.95	69.94	69.17	69.52	70.95	70.95	70.95	69.01	73.54	67.28	68.21	76.30	63.01	68.60	74.89	65.07
AT-GRU	70.66	71.21	66.47	67.97	70.66	70.66	70.66	70.00	70.07	73.55	64.74	83.24	51.45	67.27	76.09	60.54
AT-BiGRU	71.97	75.33	67.73	69.62	71.97	71.97	71.97	89.00	69.93	67.05	51.45	84.68	67.05	65.20	76.60	67.05
AT-BiLSTM	69.80	68.93	67.73	68.14	69.80	69.80	69.80	70.55	72.65	63.59	59.54	76.01	67.63	64.58	74.29	65.55
ATAE-GRU	69.94	70.11	65.51	67.11	69.94	69.94	69.94	68.97	69.73	71.64	57.80	83.24	55.49	62.89	75.89	62.54
ATAE-LSTM	68.64	68.86	65.22	66.60	68.64	68.64	68.64	69.54	68.25	68.79	60.69	78.90	56.07	64.81	73.19	61.78
ATAE-BiGRU	70.23	71.31	66.28	68.07	70.23	70.23	70.23	72.99	68.60	72.34	57.80	82.08	58.96	64.52	74.74	64.97
ATAE-BiLSTM	70.95	72.77	66.38	68.38	70.95	70.95	70.95	80.34	69.10	68.87	54.34	84.68	60.12	64.83	76.10	64.20
IAN	71.82	73.00	67.15	69.11	71.82	71.82	71.82	76.52	70.21	72.26	58.38	85.84	57.23	66.23	77.24	63.87
LCRS	68.06	67.63	64.93	65.96	68.06	68.06	68.06	70.00	69.25	63.64	56.65	77.46	60.69	62.62	73.12	62.13
CNN	67.77	66.41	64.26	65.02	67.77	67.77	67.77	66.67	70.94	61.63	53.18	78.32	61.27	59.16	74.45	61.45
GCAE	72.11	72.12	70.04	70.85	72.11	72.11	72.11	75.69	72.65	68.00	63.01	78.32	68.79	68.77	75.38	68.39
MemNet	69.65	69.09	66.76	67.68	69.65	69.65	69.65	71.17	70.57	65.52	67.05	78.32	54.91	69.05	74.25	59.75
RAM	70.09	71.32	64.93	66.48	70.09	70.09	70.09	70.62	68.84	74.51	65.32	85.55	43.93	67.87	76.29	55.27
CABASC	68.64	69.74	64.64	66.44	68.64	68.64	68.64	75.00	67.07	67.14	58.96	80.64	54.34	66.02	73.23	60.06

References

[1] Tang D, Qin B, Liu T. Aspect Level Sentiment Classification with Deep Memory Network[C]//Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. 2016: 214-224.
[2] Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural computation, 1997, 9(8): 1735-1780.
[3] Bahdanau D, Cho K, Bengio Y. Neural machine translation by jointly learning to align and translate[J]. arXiv preprint arXiv:1409.0473, 2014.
[4] Tang D, Qin B, Feng X, et al. Effective LSTMs for Target-Dependent Sentiment Classification[C]//Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers. 2016: 3298-3307.
[5] Wang Y, Huang M, Zhao L. Attention-based LSTM for aspect-level sentiment classification[C]//Proceedings of the 2016 conference on empirical methods in natural language processing. 2016: 606-615.
[6] Ma D, Li S, Zhang X, et al. Interactive attention networks for aspect-level sentiment classification[C]//Proceedings of the 26th International Joint Conference on Artificial Intelligence. AAAI Press, 2017: 4068-4074.
[7] Zheng S, Xia R. Left-Center-Right Separated Neural Network for Aspect-based Sentiment Analysis with Rotatory Attention[J]. arXiv preprint arXiv:1802.00892, 2018.
[8] LeCun Y, Bengio Y. Convolutional networks for images, speech, and time series[J]. The handbook of brain theory and neural networks, 1995, 3361(10): 1995.
[9] Xue W, Li T. Aspect Based Sentiment Analysis with Gated Convolutional Networks[C]//Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018: 2514-2523.
[10] Tang D, Qin B, Liu T. Aspect Level Sentiment Classification with Deep Memory Network[C]//Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. 2016: 214-224.
[11] Chen P, Sun Z, Bing L, et al. Recurrent attention network on memory for aspect sentiment analysis[C]//Proceedings of the 2017 conference on empirical methods in natural language processing. 2017: 452-461.
[12] Liu Q, Zhang H, Zeng Y, et al. Content attention model for aspect based sentiment analysis[C]//Proceedings of the 2018 World Wide Web Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 2018: 1023-1032.

Contacts

• jzhou@ica.stc.sh.cn

Node

Thanks for code shared by Song, Youwei, we conduct our model based on it.