Language-oriented Sentiment Analysis based on the Grammar

Structure and Improved Self-attention Network

Hien D. Nguyen

1,2,* a

, Tai Huynh

3,4,*

, Suong N. Hoang

, Vuong T. Pham

and Ivan Zelinka

Faculty of Computer Science, University of Information Technology, Ho Chi Minh City, Vietnam

Vietnam National University, Ho Chi Minh City, Vietnam

Ton Duc Thang University, Ho Chi Minh City, Vietnam

Kyanon Digital, Vietnam

Faculty of Information Technology, Sai Gon University, Ho Chi Minh City, Vietnam

Technical University of Ostrava (VŠB-TU), Czech Republic

ivan.zelinka@vsb.cz

* Equal contribution by Hien D. Nguyen and Tai Huynh

Keywords: Sentiment Analysis, Sentiment Classification, Vietnamese, Self-attention, Transformer, Natural Language

Processing.

Abstract: In the businesses, the sentiment analysis makes the brands understanding the sentiment of their customers.

They can know what people are saying, how they’re saying it, and what they mean. There are many methods

for sentiment analysis; however, they are not effective when were applied in Vietnamese language. In this

paper, a method for Vietnamese sentiment analysis is studied based on the combining between the structure

of Vietnamese language and the technique of natural language processing, self-attention with the Transformer

architecture. Based on the analysing of the structure of a sentence, the transformer is used to process the word

positions to determine the meaning of that sentence. The experimental results for Vietnamese sentiment

analysis of our method is more effectively than others. Its accuracy and F-measure are more than 91% and

its results are suitable to apply in practice for business intelligence.

1 INTRODUCTION

Sentiment analysis (SA) is one of the subfields of

Computational Linguistics and Natural Language

Processing (NLP) (Gamal et al., 2019). In the

businesses intelligence, the sentiment analysis makes

the brands understanding the sentiment of their

customers (Rokade and Kumari, 2019). They can

know what people are saying, how they’re saying it,

and what they mean. The sentiment of customer

sentiment can be found in tweets, comments, reviews,

or other places where people mention the brands.

In the current era, social network is a popular

platform for communication and interaction (Beigi,

2016). Many people found innovative information on

social network and due to that social network is the

important data source. SA is also used to detect the

https://orcid.org/0000-0002-8527-0602

https://orcid.org/0000-0002-3354-013X

https://orcid.org/0000-0002-3858-7340

influencer on the social network for the influencer

marketing (Huynh et al, 2019).

Vietnamese is a language isolate (Nguyen et al.,

2006). The meaning of a sentence belongs to the way

for organizing of its predicates (Clark, 1974). In other

words, the information about word positions

contribute the sentence meaning and grammatical

meaning. The analysing on the Vietnamese sentence

has to combine the studying of the grammar structure.

Some machine learning-based approaches have

been studied to analysis the sentiment of a

Vietnamese sentence.

CountVectorizer (Irfan et al., 2015) and Term

Frequency–Inverse Document Frequency (Tf-idf)

(Aggarwal, 2011) are used for word representations.

However, they cannot analysis the positions of words

in a sentence, so their results are not exactly. Support

Nguyen, H., Huynh, T., Hoang, S., Pham, V. and Zelinka, I.

Language-oriented Sentiment Analysis based on the Grammar Structure and Improved Self-attention Network.

DOI: 10.5220/0009358803390346

In Proceedings of the 15th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2020), pages 339-346

ISBN: 978-989-758-421-3

339

Vector Machine (Joachims, 1998) and Naïve Bayes

(Irfan et al., 2015) are used as classifiers. However,

those methods did not mention to the structure of a

sentence, so their results are not suitable in the

practice.

In (Krouska et al., 2017, Troussas et al., 2016),

authors present five well-known learning-based

classifiers (Naïve Bayes, Support Vector Machine, k-

Nearest Neighbor, Logistic Regression and C4.5) and

a lexicon-based approach (SentiStrength) to analysis

the sentiment on Twitter. However, it only studies on

English.

Besides, some types of recurrent neural networks

(RNNs), such as long short-term memory (LSTM)

(Hochreiter, 1997, Cheng et al., 2016), Bi-Directional

LSTM (biLSTM) (Schuster and Paliwal, 1997) or

gated recurrent unit (GRU) (Chung et al., 2014), are

very complex and take a long time to solve the

problem about sentiment analysis on Vietnamese.

The sentiment analysis for Vietnamese was

researched in (Nguyen et al., 2014). This study

investigated the task regarding both Support Vector

Machine (SVM) model and linguistics feature aspects

which is an annotated corpus for sentiment

classification extracted from hotel reviews in

Vietnamese. However, this method is not designed

based on the grammar structure, so some sentences

cannot be determined accurately.

Self-attention has been used successfully in a

variety of tasks including reading comprehension,

abstractive summarization, textual entailment and

learning task-independent sentence representations

(Zhou et al., 2018). The Transformer (Vaswani et al.,

2017) is the transduction model based on self-

attention to compute representations of its input and

output without using sequence aligned RNNs or

convolution. In (Hoang et al., 2019), authors study

sentiment analysis of product reviews in Vietnamese

by using Self-attention neural networks. However,

that study does not mention to the structure of

Vietnamese sentence in the analysing, so its results

are not exactly and suitable the practical

requirements.

In this paper, the method for Vietnamese

sentiment analysis is proposed. This method is used

to determine the sentiment of a sentiment sentence

including positive, negative or neutral. The structures

of a Vietnamese sentence are studied. Based on those

structures, the meaning of this sentence is analysed by

using the self-attention neural network architecture

Transformer. Besides, the layer of Squeeze and

Excitation (Hu et al., 2018) is also used to recalibrate

features in the process. The sentences will be

analysed to determine whether they are positive,

negative or neutral.

The experimental results show that our method

being more effective than other in Vietnamese

sentiment analysis. Its accuracy and F-measure are

more than 91% and its results are suitable to apply in

practice for business intelligence.

The next section presents some techniques of the

Transformer. Section 3 presents the method for

Vietnamese sentiment analysis. That method uses the

improved architecture of self-attention with

transformer on the structure of the sentences in

Vietnamese to determine their meaning. Section 4

described the experimental results. The last section

concludes the main results in this paper.

2 SELF-ATTENTION NETWORK

Scaled Dot-Product Attention: Let s

i - 1

be a query

vector q, and h

is duplicated with one is key vector k

and the other is value vector v

(in

current NLP work, the key and value vector are

frequently the same, there for h

can be considered as

or v

cav





(1)

model

exp( ) .

where , and ( , ) (2)

exp( )

jjj

eqk

aeqk









(1  j  n)

model

is the dimension of input vectors or k vector

(q, k, v have the same dimension as input embedding

vector)

Self-attention is a mechanism to apply Scaled

Dot-Product Attention to every token of the sentence

for all others.

For every token in sentence, three vectors Query,

Key, Value are created by using a linear feed-forward

layer as a transformation, then the attention

mechanism is applied to get the context matrix.

However, this process is very slow, so we consider

three matrices Q, K, V:

 Q is a matrix containing all the query vectors,

Q = [q

, q

,..., qn] with q

is a query vector.

 K is a matrix containing all the key vectors, K

= [k

, k

, ..., kn] with k

is a key vector.

 V is a matrix containing all the key vectors, V

= [v

, v

, ..., vn] with v

is a value vector.

Thus, we have:

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

340

model

(, ,) softmax . (3)

Attention Q K V V









Multi-head Attention performs the attention h times

with (Q, K, V) matrices of the dimension d

model

/h. Each

head is a time for applying Attention. For each head,

the (Q, K, V) matrices are uniquely projected with the

dimensions d

model

/h. Self-attention mechanism is

performed to yield an output of the same dimension

model

/h. After all, the outputs of h heads are

concatenated, and applied a linear projection layer

once again. The formula for this process is as follows:



( , , ) , ,..., .

ultiHead Q K V Concat head head head W



where . , . , . (4)

OOO

head QW K W V W

3 METHOD FOR VIETNAMESES

SENTIMENT ANALYSIS

In this section, the method for analysing the sentiment

of a Vietnamese sentence is proposed. The sentences

will be analysed to determine whether they are

positive, negative or neutral.

Firstly, the structures of a Vietnamese sentence

are studied. Because the scope of this study is the

evaluation comments for a product on the social

network, there are two kinds of declarative sentence

were mentioned: positive and negative sentence.

Secondly, based on those structures, the meaning

of this sentence is analysed by using the self-attention

neural network architecture Transformer. Because the

meaning of a Vietnamese sentence belongs to the

positions of words, our method is added the layer

determining the word positions into the processing

the transformer. Besides, the layer of Squeeze and

Excitation (Hu et al., 2018) is also used to recalibrate

features in the process.

3.1 Structure of a Vietnamese Sentence

Vietnamese is a language isolate. The structure of a

normal sentence of Vietnamese includes subjectum

(or thema) and praedicatum (or rhema). Subjectum is

the direct factor of a sentence describing the scope of

thing which is mentioned in the second direct factor -

praedicatum (Cao, 2017).

There are three frequent sentence types:

declarative, interrogative, and imperative. The

declarative is subject to judgments of truth and

falsehood (Cao, 2017). The interrogative elicits a

verbal response from the addressee. The imperative

indicates the speaker’s desire to influence future

events. In the problem about sentiment analysis, we

only need to determine whether a sentence is positive,

negative or neutral; thus, in the scope of this paper,

we only mention to the declarative sentence type.

The structure of a single declarative sentence in

Vietnamese is shown in Fig.1:

Figure 1: Structure of a single declarative sentence in

Vietnamese.

Definition 1: Kinds of the structure of a positive

sentence

A single positive declarative sentence in

Vietnamese has the foundation structure:

It is classified as Table 1.

Table 1: Kinds of the structure of a positive sentence.

Kinds Variants

P is <noun>:

= <S><noun>

<noun>

P is <verb>:

= <S> <verb>

><verb><S

><P

><verb><P

><object>

P is <adjective>:

= <S> <adj.>

P is <noun><adj.> with <noun> belongs to <S>

Definition 2: Kinds of the structure of a negative

sentence

A single negative declarative sentence in

Vietnamese has the foundation structure:

It is classified as Table 2.

Language-oriented Sentiment Analysis based on the Grammar Structure and Improved Self-attention Network

341

Table 2: Kinds of the structure of a negative sentence.

Kinds Variants

P is <noun>:

= <S> <negative

word> <noun>

<S><negative word> “là”

<noun>

<S> <negative word> <word

of kind> <noun>

P is <verb>:

= <S> <negative

word><verb>

><negative word> <verb>

><P

> <negative word> <verb>

><object>

P is <adjective>:

P is <noun><adj.> with <noun> belongs to <S>

In a Vietnamese declarative sentence, each word

has to been appeared orderly. Although two sentences

have the same referent, “same referent” means they

both describes an objectivity fact, they are not identity

about the meaning. The meaning of a sentence

belongs to the way for organizing of its predicates. In

other words, the information about word positions

contribute the sentence meaning and grammatical

meaning.

Some characteristics of an isolate language,

especially Vietnamese, for learning context are as

follows:

●

In linguistic activities, words do not change

their morphemes. Grammatical meanings are

not included in words.

●

Formal word, word position and word order

clarify the grammatical relationship as well as

the grammatical meaning of words and

sentences. Example: Add the formal words “sẽ”

(will) or “đang” (_ing) before “học” (study) will

change the tense of the action. Another example

of reversing words also changes the meaning of

grammar, for example "chân bàn" (leg of table)

and "bàn chân" (foot).

●

The lines between syllables, morphemes and

words are not clear. Example: In Vietnamese

"nhà" is a morpheme, and also is a word.

The main point of this research is around the

importance of word position information to

contribute sentence meanings and grammatical

meanings.

3.2 Pre-processing Method

Datasets will be gone through a pre-processing

pipeline of the text documents. Some available

research, such as sentence segmentation, normalize

the text, word segmentation and noise cleaning, were

mainly used to do this pipeline automatically.

Sentence segmentation is a procedure to split a

paragraph into sentences. Then, each sentence will be

text normalized.

In the text normalization, the input will be low

cased. Next, all the links, phone numbers and email

addresses were replaced by “urlObj”,

“phonenumObj” and “mailObj”, respectively.

Finally, words tokenizer from Underthesea (2019) for

Vietnamese was also applied. The input text will be

split into words, phrases, or other meaningful parts,

namely tokens.

3.3 Word Embedding

The fastText (2019) is used for word embedding. In

many cases, users may type a wrong word

accidentally or intentionally. fastText deals with this

problem very well by encoding at the character level.

In case having a wrong word, very rare words or out-

of-vocabulary words, fastText still can represent them

with an embedding vector that most similar to word

met in trained sentences.

There had been no fastText pre-trained model for

Vietnamese spoken language. Therefore, we trained

fastText for Vietnamese vocabulary as embedding

pre-trained weights from a corpus over 70,000

documents of multi-products reviews crawled from e-

commerce sites mentioned above with no label. Rare

words that occur less than 5 times in the vocabulary

were removed. Embedding size is 384. After training,

we have 5,534 vocabularies in total.

3.4 Sentiment Analysis in Vietnamese

In original architecture of Transformer, the position

encoding for a word is summed with Context

encoding from pre-trained fastText model (with same

dimensions of features). After this process, the

outputs were applied a linear projection to create

three vectors Q (query), K (key), V (value) as input

for Multi-head Attention layer:

.( )

WCP





(5)

where A is one of the three vectors Q, K, or V,

inputs of Multi-head Attention, which were

mentioned in Section 2. C is the context encoding

with

model

dimension, and P is the position encoding

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

342

Figure 2: The process of Vietnamese sentiment analysis.

with d

model

dimension too. W

is a matrix of trainable

weights with the size d

model

 d

model

From (5), we have:

AWCWP

(6)

That means, Context information and Position

Information both play the same role to create

semantic meaning of a word in the sentence.

Nonetheless, in Vietnamese, the meaning of a

sentence belongs to the information of word

positions. Thus, the concatenate operator is used to

combine the information of word positions and the

inputs of Multi-head Attention layer. That makes the

context information and position information having

the different weights during the transformation

process.

AW X

(7)

(,)X Concatenate C P

(8)

The dimension of X is 2*d

model

, then W

’ is a

matrix of trainable weights with the size 2*d

model



2*d

model

; thus, P and C are not the same weight as (6).

It means: if the meaning belongs to the word order,

the weight of the position (P) will be larger; else if it

is the formal word, the weight of the context (C) will

be larger.

The process for the Vietnamese sentiment

analysis is shown in Fig.2. The proposed model for

sentiment analysis in Vietnamese is based on Self-

attention with Transformer architecture. In this work,

the “concatenate” operation is used to incorporate

position information with word context information

as the input to Multi-head Attention layer as (7)(8).

The input sentence is transformed to context

embedding by using self-attention mechanism.

Moreover, in Fig.2, the layer of Squeeze and

Excitation (Hu et al., 2018) is added between Multi-

head Attention and Feed Forward layer to recalibrate

features. It uses global information from the context

matrix, which was the result of the Multi-head

Attention layer, to select important features and

suppress less useful ones before performing a

transformation with feed forward network layer. It

helps network to learn more important features

efficiently for the task of sentiment analysis.

The gating mechanism of the Squeeze and

Excitation (SE) layer is performed by stacking a

GobalAveragePooling1D layer then forming a

bottleneck with two dense layers. The first layer is a

dimensionality-reduction layer with reduction ratio r

with a non-linear activation. The second layer is the

dimensionality-increasing layer to return the result of

a sigmoid activation function with the dimension

model

iij





(9)

where, f

is the i

feature of the context matrix F,

j is the j

token of the sequence.

The output of this process is the squeeze global

information of a feature of sequence into a feature

channel descriptor. Then, every feature in the context

matrix are represented by a value of this descriptor.

After that, a bottle-neck dense net is used to select

the important features for sentiment task efficiently.

(.(.))

SWWX





(10)

where, δ refers to the ReLU function, S is the

feature channel descriptor, σ is a sigmoid activation

function, and W

fc2

, W

fc1

are trainable weights of the

network.

The final output of this layer is obtained by the

feature-wise multiplication between the scalar S and

the context matrix:

scale

(,)OM FS

(11)

where, O is the recalibration of the context matrix,

and M

scale

(F, S) refers to the feature-wise

multiplication between the scalar S and the context

matrix F.

The work of the SE layer is shown in Fig. 3:

Language-oriented Sentiment Analysis based on the Grammar Structure and Improved Self-attention Network

343

Figure 3: The layer of Squeeze and Excitation.

4 EXPERIMENTAL RESULS

4.1 Dataset

Dataset is set of comments of electronic products

which were crawled from Vietnamese e-commerce

websites, such as Tiki, Lazada, Shopee, Sendo,

Adayroi, Dienmayxanh, Thegioididong, FPTShop,

Vatgia. It includes 32,953 documents in labelled

corpus: 22,335 positives documents and 10,618 of

negatives documents.

For making the dataset balanced, some short

negative documents are duplicated and segmented the

longer ones. In the final result we have over 43, 500

documents in corpus with 22, 335 positives and 21,

236 negatives. Using for training models, we splitted

corpus into 3 sets as following:



Training set: 27, 489 documents.



Validation set: 6, 873 documents.



Test set: 8, 591 documents.

4.2 Evaluation Measures

Four measures which have been used in this study are

based on the confusion matrix output. They are True

Positive (TP), False Positive (FP), True Negative

(TN), and False Negative (FN).

Precision (P) = TP/(TP + FP)

Recall (R) = TP/(TP + FN)

Accuracy (A) = (TP + TN)/(TP + TN + FP + FN)

F-measure = 2.(P.R)/(P + R)

4.3 Results of Testing

We compare our model with four base line RNN

models such as Long-Short Term Memory (LSTM),

Gated Recurrent Units (GRU), stacked bidirectional

LSTM and stacked bidirectional GRU with the

following configurations. All models were evaluated

on 8,591 documents.



LSTM and GRU: 1 layer with 1,200 units.



Stacked bidirectional model of LSTM and

GRU: 2 stacked layers with 1,200 units in

forward and 1,200 units in backward for

each layer.

These model are compared with our method: i/

improved Self-Attention without the SE layer, ii/

improved Self-Attention combining the SE layer. The

comparison results based on evaluate measures are

shown in Table 3.

The method using the improved Self-attention

with SE layer is more effective than other for

Vietnamese sentiment analysis. Besides, this method

is also more useful about the inference time.

From the experimental results, the improved Self-

attention has the accuracy and F-measure is better

than original methods of GRU and LSTM. Although

the precision and F-measure of the improved Self-

attention are lower than improved methods of GRU

and LSTM (Stacked bi-GRU and Stacked bi-LSTM,

its inference time is faster because it worked based on

the grammatical structure of Vietnamese sentence.

The SE layer helps to select important features and

remove less useful ones before performing a

transformation with feed forward network layer. It

helps network to learn more efficiently for the task of

sentiment analysis. From that, the improved Self-

Table 3: The results of the comparison between methods.

(%)

F-measure

(%)

Time

(s)

GRU 58.9 58.5 58.7 58.7 0.25

Stacked bi-GRU 91.1 90.8 90.9 91 1.05

LSTM 63.7 61.7 61.3 62.7 0.38

Stacked bi-LSTM 89.6 89.2 89.3 89.4 1.63

Improved Self-attention 86.6 85.1 85.3 85.9 0.003

Improved Self-attention

combining SE layer

91.7 91.6 91.6 91.6 0.07

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

344

attention combining SE layer is more precise than

other methods, it has the accuracy and F-measure

being more than 91%. Moreover, the inference time

of the combining method is also better than others.

Hence, the proposed method can be useful in practice,

especially in business intelligence.

5 CONCLUSIONS

In this paper, a method for sentiment analysis in

Vietnamese is proposed. This method is studied based

on the combination between the structure of a

Vietnamese sentence and the technique of NLP, the

self-attention with Transformer. The structures of a

declarative sentence are studied and applied in the

analysing of their meaning. Based on those structures

of the sentences, the Self-attention network with the

Transformer is used to analysis the sentiment of the

sentence. The Self-attention network is improved by

two steps:

(1) Adding the layer to determine the word

positions by using the formulas (7)(8).

(2) Adding the layer of Squeeze and Excitation

between Multi-head Attention and Feed

forward layer to recalibrate features.

The experimental results of our method for

Vietnamese sentiment analysis has the accuracy more

than 91%, it is more effective than other methods.

Besides, the inference time of the proposed method is

also better than others. The process of this method can

be applied in business for analysing the information

on social network which serves in the influencer

marketing.

In practice, the vast amount of training examples

necessary to get satisfactory results is an obstacle to

develop the natural language processing. In the

future, we will use the method to transform this paper

proposes a method for transforming the sentiment of

a sentence to the opposite sentiment (Leeftink and

Spanakis, 2019). This method can reduce by half the

work required in the generation of training examples.

In the real-word, people can show their views in a

sarcastic way that is difficult to determine. In the

future work, the method need to be developed to

classify the sentiment in those cases. That

improvement has to analysis deeper in the sentence’s

structure and the technique of self-attention network.

Moreover, for applying in business intelligence, such

as the influencer marketing, the sentiment analysis in

Vietnamese will be used to design the method for

detecting the influencer on the social network, which

were presented by the relational model (Do et al.,

2018, Nguyen et al., 2015).

ACKNOWLEDGEMENTS

This research is supported by VinTech Fund, a grant

for applied research managed by VinTech City, under

grant number DA132-15062019.

REFERENCES

Aggarwal, C., Wang, H. 2011. Text Mining in Social

Networks. In Social Network Data Analytics,

Springer, 353-378.

Beigi, G., Hu, X., Maciejewski, R., Liu, H. 2016. An

Overview of Sentiment Analysis in Social Media and

Its Applications in Disaster Relief. In Sentiment

Analysis and Ontology Engineering: An Environment

of Computational Intelligence. Springer, 313-340.

Cao, H. 2017. Vietnamese: Review of Functional Grammar,

Publisher of Social Science, 2

edition. (Vietnamese)

Cheng, J., Dong, L., Lapata, M. 2016. Long short-term

memory-networks for machine reading. In EMNLP

2016, Conference on Empirical Methods in Natural

Language Processing, Nov. 2016.

Chung, J., Gulcehre, C., Cho, K., Bengio, Y.

2014. Empirical evaluation of gated recurrent neural

networks on sequence modeling. In NIPS 2014

Workshop on Deep Learning, Dec. 2014.

Clark, M. 1974. Passive ane Ergative in Vietnamese. In:

Nguyen Dang Liem (ed.), South-East Asian Linguistic

Studies, 75-80.

Do, N., Nguyen, H., Selamat, A. 2018. Knowledge-Based

model of Expert Systems using Rela-model.

International Journal of Software Engineering and

Knowledge Engineering (IJSEKE) 28(8), 1047 – 1090.

fastText. 2019. https://fasttext.cc/

Gamal, D., Alfonse, M., El-Horbaty, E.M., Salem, A.M.,

2019. Implementation of Machine Learning Algorithms

in Arabic Sentiment Analysis Using N-gram Features.

Procedia Computer Science 154(2019), 332-340.

Hoang, S., Nguyen, L., Huynh, T., Pham, V. 2019. An

Efficient Model for Sentiment Analysis of Electronic

Product Reviews in Vietnamese. In FDSE 2019, 6

International Conference on Future Data and Security

Engineering. Springer. LNCS 11814, 132–142.

Hochreiter, S., Schmidhuber, J. 1997. Long short-term

memory. In Neural Computing 9(8), 1735–1780.

Hu, J., Shen, L., Sun, G. 2018. Squeeze-and-Excitation

Networks. In CVPR 2018, Conference on Computer

Vision and Pattern Recognition, June 2018.

Huynh, T., Zelinka, I., Pham, Nguyen, H. 2019. Some

measures to Detect the Influencer on Social Network

Based on Information Propagation. In WIMS 2019, 9

International Conference on Web Intelligence, Mining

and Semantics, June 2019.

Irfan, R., et al. 2015. A survey on text mining in social

networks. In The Knowledge Engineering Review

30(2), 157-170.

Language-oriented Sentiment Analysis based on the Grammar Structure and Improved Self-attention Network

345

Joachims, T. 1998. Text Categorization with Support

Vector Machines: Learning with Many Relevant

Features. In ECML 1998, 10th European Conference

on Machine Learning Chemnitz, April 1998. Springer,

137-142.

Krouska, A., Troussas, C., Virvou, M. 2017. Comparative

Evaluation of Algorithms for Sentiment Analysis over

Social Networking Services. Journal of Universal

Computer Science 23(8), 755-768.

Leeftink, W., Spanakis, G. 2019. Towards Controlled

Transformation of Sentiment in Sentences. In ICAART

2019, 11

International Conference on Agents and

Artificial Intelligence, Feb. 2019.

Nguyen, B., Nguyen, H., Romary, L., Vu, L. 2006. Lexical

descriptions for Vietnamese language processing. In

Language Resources and Evaluation 40(3/4), Asian

Language Processing: State-of-the-Art Resources and

Processing, 291-309.

Nguyen, D., Ngo, B., Tu, P. 2014. An empirical study on

sentiment analysis for Vietnamese. In ATC 2014, 7

International Conference on Advanced Technologies

for Communications, Oct. 2014.

Nguyen, H., Pham, V., Le, T., Tran, D. 2015. A

Mathematical Approach for Representation Knowledge

about Relations and Its Application. In KSE 2015, 6

International Conference on Knowledge and Systems

Engineering, Oct. 2015.

Rokade, P., Kumari, D.A. 2019. Business intelligence

analytics using sentiment analysis - a survey.

International Journal of Electrical and Computer

Engineering (IJECE) 9(1), 613-620.

Schuster, M., Paliwal, K. 1997. Bidirectional recurrent

neural networks. IEEE Trans. Sig. Process. 45(11),

2673–2681.

Troussas, C., Krouska, A., Virvou, M. 2016. Evaluation of

ensemble-based sentiment classifiers for Twitter data.

In IISA 2016, 7

International Conference on

Information, Intelligence, Systems & Applications, July

2016.

Underthesea. 2019. http://undertheseanlp.com/#!/

Vaswani, A., et al. 2017. Attention is all you need. In NIPS

2017, 31

Conference on Neural Information

Processing Systems, Dec. 2017.

Zhou, Y., Zhou, J., Liu, L., Feng, J., Peng, H., Zheng, X.

2018. RNN-based sequence-preserved attention for

dependency parsing. In AAAI 2018, 32

AAAI

Conference on Artificial Intelligence, Feb. 2018.

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

346