CNN-LSTM-CRF for Aspect-Based Sentiment Analysis:

A Joint Method Applied to French Reviews

Bamba Kane, Ali Jrad, Abderrahman Essebbar, Oph

elie Guinaudeau, Valeria Chiesa, Ilhem Qu

enel

and St

ephane Chau

Research and Innovation Direction, ALTRAN Sophia-Antipolis, France

bamba.kane, ali.jrad, abderrahman.essebbar, ophelie.guinaudeau, valeria.chiesa, ilhem.quenel,

Keywords:

Natural Language Processing (NLP), Aspect-Based Sentiment Analysis (ABSA), Convolutional Neural

Network (CNN), Long Short-Term Memory (LSTM), Conditional Random Field (CRF), SemEval.

Abstract:

Aspect Based Sentiment Analysis (ABSA) aims to detect the different aspects addressed in a text and the

sentiment associated to each of them. There exists a lot of work on this topic for the English language, but

only few models are adapted for French ABSA. In this paper, we propose a new model for ABSA, named CLC,

which combines CNN (Convolutional Neural Network), Bidirectional LSTM (Long Short-Term Memory) and

CRF (Conditional Random Field). We demonstrate herein its great performance on the SemEval2016 French

dataset. We prove that our CLC model outperforms the state-of-the-art models for French ABSA. We also

prove that CLC is well adapted for other languages such as English. One main strength of CLC is its ability

to detect the aspects and the associated sentiments in a joint manner, unlike the state-of-the-art models which

detect them separately.

1 INTRODUCTION

Over the last few years a lot of work has been done

in the ﬁeld of sentiment analysis (Lin and Luo, 2020).

Most of the work focuses on identifying the sentiment

that is expressed in a text, in a global way by giv-

ing a positive, a negative or a neutral appreciation to

the whole text. However, such a task is sometimes

too coarse. For example, given the following review

on a restaurant, ”the meat was delicious but the price

was exaggerated”, it is quite difﬁcult to highlight an

overall sentiment. Aspect Based Sentiment Analysis

(ABSA) is a sub-domain of Natural Langage Process-

ing (NLP) (Thet et al., 2010) which takes up this chal-

lenge and provides a much ﬁner analysis by identify-

ing the sentiment associated to each aspect evoked in

the text.

In the previous example, ABSA objective is to

detect both the two categories of aspects addressed

in the text (food

quality and food price), and the

sentiment associated to each aspect (positive for

food quality and negative for food price). A word

(or a group of words) denoting an aspect is called

target. In this review, ”the meat was delicious but

the price was exaggerated”, the target for the aspect

food quality is meat, and the one for food price is

price.

ABSA can be split into three tasks (Pontiki et al.,

2016):

• Opinion Target Extraction (OTE): aims to detect

the words that constitute the targets.

• Aspect Category Detection (ACD): focuses on the

detection of the different types of aspects that are

evoked in a text.

• Aspect Sentiment Polarity (ASP): assigns a sen-

timent polarity (positive, negative or neutral) for

each identiﬁed aspect.

In this paper, we focus on ACD.

There is an increased interest in ABSA especially

with the dedicated challenges of SemEval (Pontiki

et al., 2016) that offers a framework to deﬁne and

evaluate models on different domains. It was initiated

for English language and it is now open to several

other languages such as French or Spanish (Pontiki

et al., 2016).

In order to detect the aspects and the related sen-

timents, many methods have been proposed. In (Kir-

itchenko et al., 2014) the authors use Support Vec-

tor Machine (SVM) for this purpose, while (Hamdan

et al., 2015) use Conditional Random Field (CRF).

Convolutional Neural Networks (CNN) have also

498

Kane, B., Jrad, A., Essebbar, A., Guinaudeau, O., Chiesa, V., Quénel, I. and Chau, S.

CNN-LSTM-CRF for Aspect-Based Sentiment Analysis: A Joint Method Applied to French Reviews.

DOI: 10.5220/0010382604980505

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 1, pages 498-505

ISBN: 978-989-758-484-8

been widely employed for ABSA (Ruder et al., 2016;

Wu et al., 2016).

However, most of these works focus on English

language and there is only a few work in the litera-

ture for the French language. The lack of linguistic

tools or resources is probably the explanation. Since

the French language has it own characteristics such as

gender for nouns, liaisons, or prepositions which are

very different from English, the existing resources are

not suitable.

In this article, we propose a new model named

CLC which combines a CNN to encode word-level

information into its word-level representation and a

bidirectional Long Short Term Memory (BiLSTM) to

model context information of each word. Finally, we

use a sequential CRF to get the best sequence of as-

pects and the related sentiments. Our proposed ap-

proach leads to three main contributions:

• a new architecture for ABSA, named CLC, allow-

ing to detect the aspects and associated sentiment

in a joint manner;

• better performance than state-of-the-art methods

for ABSA on the French language;

• great performance on English SemEval2016

dataset.

The paper is organized as follows: in section 2, we de-

scribe the state-of-the-art methods for ABSA. In sec-

tion 3, the new model CLC is detailed. In section 4,

the experimentation details are given. We show and

analyze the performance on the SemEval2016 French

dataset about restaurants and we provide a compari-

son with the state-of-the-art methods.

2 STATE OF THE ART

2.1 Aspect Based Sentiment Analysis

(ABSA)

In this section, we present the state-of-the-art meth-

ods for ABSA and then we detail the methods that are

relevant for French language.

2.1.1 Traditional Methods for ABSA

• NRC-Canada: (Kiritchenko et al., 2014) de-

ployed a traditional Support Vector Machine

(SVM)-based model and added some extensive

feature engineering like Part of Speech (POS)

tags, various types of n-grams, and lexicon fea-

tures.

• Rec-NN (Recursive Neural Networks): (Dong

et al., 2014) used, in a ﬁrst step, rules in order

to get the word related to the aspect at the top of a

dependency tree. Then, the representation of the

sentence towards the aspect was learned by a se-

mantic composition using Recursive Neural Net-

work.

• DDG (Domain Dependency Graph): (Kumar

et al., 2016) used information extracted from de-

pendency graphs learned on different domains and

languages of SemEval2016 and showed very efﬁ-

cient results on different languages (among which

the French language).

The two followings methods participate to the Se-

mEval2016 challenge for French restaurant reviews:

• XRCE: it was the challenge winner for French

language (Brun et al., 2016; Brun and Nikoulina,

2018). Their method was composed by two steps

: 1) Classiﬁcation process is performed using

the output of a Conditional Random Field (CRF)

which has been specialized at word level on the

available training data (labeled aspects and senti-

ments) in order to classify terms into one or more

aspect categories. 2) At sentence level, the classi-

ﬁcation models associate aspect categories of sen-

tences with probabilities. The aspect categories

are then assigned using a threshold over the as-

signed probabilities

• BUTknot (Machacek, 2016): this method got

the second rank for the SemEval2016 challenge

in French. The aspect detection is based on su-

pervised machine learning using bi-grams bag-of-

words model for multi-languages. The perfor-

mance is enhanced by a term substitution tech-

nique. The system has reached a very good per-

formance in comparison with other submitted sys-

tems for the challenge.

2.1.2 Deep Learning Methods for ABSA

Many deep learning architectures are used for ABSA

(Young et al., 2018):

• CNN (Ruder et al., 2016): proposed a method us-

ing multiple ﬁlters and obtained competitive re-

sults on both polarity and aspect detection tasks.

(Poria et al., 2016) designed a seven-layer CNN

architecture and made use of both POS tagging

and word embeddings as features.

• MEM-Net (Memory Network): uses the inter-

aspect relations modeling (Tang et al., 2016b).

This method was used on French SemEval2016

datasets.

CNN-LSTM-CRF for Aspect-Based Sentiment Analysis: A Joint Method Applied to French Reviews

499

LSTM for ABSA. The LSTM model has achieved

many great successes in various NLP tasks and par-

ticularly for ABSA. Following, we present different

models for ABSA using LSTM (Tang et al., 2016a):

• TD-LSTM (Target-Dependent LSTM): it deﬁnes

the use of LSTM by considering the target. The

BiLSTM network is used to model both the left

and the right context of a target. Then the left and

right target-dependent representations are con-

catenated for predicting the sentiment polarity of

the target.

• TC-LSTM (Target-Connection LSTM): it uses a

BiLSTM for which semantic relatedness of target

with its context words are incorporated. A target

vector is calculated by averaging the vectors of the

words that compose the target. This method has

optimal performance by using a simple average of

the word embeddings of a target.

• BiLSTM-CRF (Chen et al., 2017): it is a combi-

nation of BiLSTM and CRF layers to extract the

targets related to the aspects. (Kooli and Pigneul,

2018) also uses BiLSTM and CRF for aspect ex-

traction but only on French datasets.

Attention Mecanism. Opposite polarities may be

detected when multiple aspects are considered. Thus,

the use of only target information is not sufﬁcient.

The application of an attention mechanism can extract

the association of important words denoting an aspect

(Wang et al., 2016). It allows to capture the key part

of a sentence in relation with a given aspect and to

identify the associated sentiment.

Let us consider the following example: ”The

menu is great but service is a disaster”. For the tar-

get menu, the word great will have a high weight and

for the target service, the word disaster will have a

high weight. Through attention mechanism, the net-

work can learn the association of a positive sentiment

for the menu and and a negative one for the service.

• Attention-based LSTM with Aspect Embed-

ding (ATAE-LSTM): it strengthens the effect of

target embeddings by appending it with each word

embeddings and uses BiLSTM with attention to

get the ﬁnal representation for the detection of the

aspects and the related sentiments. (Wang et al.,

2016).

• Interactive Attention Network (IAN): (Ma

et al., 2017) uses this model to learn the represen-

tations of the target and its context with BiLSTMs

and attentions interactively, that generates the rep-

resentations for targets and contexts with respect

to each other.

• Attentional Encoder Network (AEN): (Song

et al., 2019) do not use recurrence and employ at-

tention based encoders for modeling context and

aspect.

3 CLC: A NEW ARCHITECTURE

FOR ABSA

In this section, we present our new model: CLC. It

combines CNN, BiLSTM and CRF layers. Here, we

describe each layer of the proposed neural network.

We also motivate our approach and explain why each

layer of the model is used. The global architecture of

the proposed model is illustrated in Figure 4.

3.1 Embedding Layer

Embedding is used to represent each word in a vector

with deﬁned dimensions. The resulting vectors repre-

sent the projections in a continuous vector space. The

position of a word is calculated as a function of the

words in its context. The embedding for a word cor-

responds to the position of this word in the learned

vector space. Nowadays, there are a lot of pre-trained

word embeddings on corpus of millions of words like

Word2Vec, GloVe or Wikipedia2Vec

3.2 CNN

We use CNN layer in order to get a word-level rep-

resentation for each word of a phrase. (Kim, 2014)

proved that a simple CNN gives very good perfor-

mance for text classiﬁcation with one layer of con-

volution on top of word vectors from an unsupervised

neural language model. (Chiu and Nichols, 2016) also

demonstrated that CNN is efﬁcient for taking into ac-

count some morphological information like sufﬁx or

preﬁx of a word and encodes it into a neural represen-

tation. Figure 1 shows the different steps we used for

the CNN layer in the CLC model.

3.3 LSTM

LSTM has been proposed to solve the vanish-

ing/exploding gradient problem related to Recurrent

Neural Network (RNN). A LSTM unit is composed of

three gates: input, output and forget which can check

the proportion of information to forget or to keep for

the next step. Figure 2 shows in details a LSTM unit

with its different components.

The formulas for updating a LSTM unit at time t

are:

NLPinAI 2021 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

500

Figure 1: CNN on text (Nguyen et al. 2019).

Figure 2: LSTM unit (Yuan et al. 2019).

= σ(W

t−1

+ b

)

= σ(W

t−1

+ b

)

= tan(W

ht − 1 +U

+ b

)

= f

t−1

+ i

= σ(W

t−1

+ b

)

= o

tan(c

)

where σ is the element-wise sigmo

ıd function and

is the element-wise product. x

corresponds to the in-

put vector (like word embedding) at time t, and h

the hidden state (or output) vector storing all the use-

ful information at (and before) time t. U

, U

denote the weight matrices of different gates for in-

put x

, and W

, W

are the weight matrices for

hidden state h

. b

, b

denote the bias vectors.

BiLSTM consists of two LSTM: one with the in-

put in a forward direction, and the other in a back-

wards direction. BiLSTM allows to fully take into

account the context of each word. Thus, we can learn

the information from context which facilitates the as-

pect detection and the analysis of related sentiments.

Figure 3 shows the BiLSTM architecture.

Figure 3: BiLSTM architecture (from S.Cornegruta and al,

2016.

3.4 CRF

CRF is used in statistical modelling for sequence la-

beling, which can be applied for aspect detection and

ABSA.

The CRF layer with its constraints ensure that the

predicted labels are valid. During the training, these

constraints are learned automatically. Formally, let

z = {z

, ..., z

} be a phrase with z

, the vector for the

word and y = {y

, ..., y

} be the generic sequence

of aspects for z. γ(z) constitutes the set of possible

aspects for z. CRF corresponds to a family of condi-

tional probability P(y|z;W ; b) over all possible aspect

sequences y given z and can be formulated as follows:

P(y|z;W ;b) =

∏

i=1

i−1

, y

, z)

∑

∈γ(z)

∏

i=1

i−1

, y

, z)

where ψ

, y, z) = exp(W

+ b

) are potential

functions, and W

and b

are the weight vector and

bias corresponding respectively to label pair (y

, y).

3.4.1 IOB Tagging

IOB (Inside, Outside, Begin) is a widely used tagging

scheme. The B- tag corresponds to the beginning of

an annotated chunk (target). The I- tag indicates that

the tag is inside a chunk. Only a B- tag (or another I-

tag for n-grams chunks) could precede an I- tag. The

O- tag corresponds to the tokens that do not belong to

any annotated chunks.

We can deﬁne the constraints of the CRF layer by

using the IOB scheme. Hence, we can deﬁne some

constraints such as:

• the ﬁrst word of a sentence starts with a B- or an

O- tag, and never with an I- tag;

• an O- tag followed by an I- tag is an invalid se-

quence;

• the ﬁrst tag for a target is a B- tag and never an I-

tag.

CNN-LSTM-CRF for Aspect-Based Sentiment Analysis: A Joint Method Applied to French Reviews

501

Figure 4: Architecture of CLC. We observe that only the word w

denotes an aspect. At the CRF layer, we predict a

as the

aspect and s

as the associated sentiment.

Table 1: IOB Tagging.

meat price is affordable

B Meal#Price I Meal#Price O O

For example, if the sentence ”meat price is afford-

able” is encoded with the IOB tagging scheme, we

can have this labeling:

We notice in Table 1 that the target meat price is

annotated with a B- and an I- tag for the aspect

Meal#Price. It means that we consider here an unique

target composed of two words and not two distinct tar-

gets (for the aspect Meal#Price).

3.5 Joint Aspect and Sentiment

Detection

Many ABSA models used separate methods for the

two following tasks:

• detection : detecting all the aspects;

• classiﬁcation : classifying the polarity for each

detected aspect.

In this paper, we propose a joint model able to detect

the aspects and the associated sentiments at the same

time.

Our CLC model is described in Figure 4. As men-

tioned earlier, the CNN layer is an effective approach

to extract morphological information like the preﬁx or

sufﬁx of a word and encode it into neural representa-

tion. This allows a better word-level representation.

The input used for the CNN layer is the word em-

beddings. Then, the CNN layer acts as a character

embedding. We feed also a BiLSTM layer with the

word embeddings. The output vectors from BiLSTM

and CNN are merged and ﬁnally, the merged vectors

are given as input to the CRF layer in order to predict

the aspects and the related sentiments.

Let W and T be the two inputs that we used for

ﬁtting the model: W corresponds to the words and T

to the corresponding tags. In Table 2 we can see an

example of W and T given a customer review.

Table 2: Inputs for an example of a customer review.

W Price was affordable

T B Meal#Price O O

In order to jointly detect aspects and sentiments, we

insert an index for polarity just after the correspond-

ing aspect in W and T and we give these new inputs

to the model for the training (see Table 3). This al-

lows CLC to detect aspects and sentiments at the same

time.

Table 3: Inputs for joint aspect and sentiment detection.

W Price [sentiment] was affordable

T B Meal#Price positive O O

For these new inputs dedicated to joint detection, we

also need to encode the sentiments. Table 4 shows

the encoding we use for positive, negative or neutral

sentiment. If there is n aspects in the dataset, we can

NLPinAI 2021 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

502

Figure 5: Example of review from French SemEval2016 dataset on restaurants. We note that the aspect is tagged by category.

For implicit aspect the target is NULL, because no words in the sentence is directly referring to the aspect.

observe in table 4 that the positive sentiment is en-

coded by n+1, the negative sentiment is encoded by

n+2 and the neutral one by n+3. It is an arbitrary

choice in order to avoid confusion with the encoding

of the vocabulary words.

Table 4: Encoding of the sentiments for joint detection with

n being the number of aspects.

Sentiment Encoding

positive n + 1

negative n + 2

neutral n + 3

4 EXPERIMENT

In this section, we show how our CLC model is well

adapted for aspect and sentiment detection on French

language. We give details about the dataset we used

and we demonstrate that CLC outperforms the exist-

ing methods for French ABSA.

4.1 SemEval2016 Dataset

French Dataset. In our experiment, we use the

French SemEval2016 annotated dataset about restau-

rants (Apidianaki et al., 2016). It contains 337 re-

views for the training set and 120 reviews for the test

set. The dataset is annotated with targets, aspects and

sentiments (Figure 5). Since each review can be com-

posed of several sentences, the dataset actually con-

tains 1669 annotated sentences for training set and

696 for the test set. It incorporates 12 categories of

aspects and 3 types of sentiments (positive, negative

and neutral). In Figure 5, we give an extract of the

dataset.

English Dataset. In order to conﬁrm that CLC is

adapted for other languages, we also test our model

with a SemEval2016 annotated dataset in English. In

the training set of this dataset, there are 2000 sen-

tences coming from 350 reviews. The test set contains

676 sentences from 90 reviews.

4.2 Training Details

Before training our model we perform some pre-

processing steps such as removing the stop words and

doing lemmatization in order to improve the CLC per-

formance. We use publicly available Wikipedia2Vec

with 300-dimensional embeddings trained on words

from Wikipedia.

For CNN Layer, we used 50 ﬁlters with window

length equal to 4. The BiLSTM layer network size is

300.

4.3 Results

Table 5 shows a comparison of our CLC method with

the state-of-the-art models. Since we are interested

in French language, we only consider methods that

give the highest performances on the French SemEval

datasets. We present the F-score for aspect detection

using the French SemEval2016 dataset on restaurants.

Generally, the reference methods only focus on aspect

detection or give separate results for the aspects and

the sentiments.

We can observe in Table 5, that our CLC model

outperforms the state-of-the-art models when we fo-

cus on ACD. Therefore, CLC shows great interest

when it comes to the French language. Thus, CLC

can be used to ﬁll the gap that exists about efﬁcient

methods for aspect detection in French language.

During our study, we observed that using a CRF

layer instead of a softmax activation improves the F-

score from 74.2% to 77.2%. The reason is that the

CRF layer allows to predict the label using neighbour-

ing tagging predictions that eliminates some false pre-

dictions.

As mentioned in section 3.5, our CLC model can

CNN-LSTM-CRF for Aspect-Based Sentiment Analysis: A Joint Method Applied to French Reviews

503

Table 5: Aspect detection Comparison for French SemEval-2016 dataset about restaurants.

F1-Score

(Kumar et al., 2016) 57.8

(Ruder et al., 2016) 61.2

(Brun et al., 2016) 61.2

(Poria et al., 2016) 67.06

(Chen et al., 2017) 73.50

(Kooli and Pigneul, 2018) 70.97

CLC 77.2

also be deployed for joint aspect and sentiment de-

tection. We observe that when we use CLC as a joint

model, the F1-score for the couple (aspect, sentiment)

reaches 88.7%. The reason is that CLC learns well

the correlation between the aspect and the sentiment.

Thus, more aspects are well detected.

CLC Applied on Other Datasets. In order to prove

the generalization capability of CLC, we apply it on

an unseen dataset which is the french SemEval2016

dataset about Museum. The Museum dataset contains

162 french annotated reviews with 668 sentences for

testing. CLC achieves great performances in this con-

text with a a F1-score equal to 69.97%

To prove that CLC is also adapted for English lan-

guage, we tested it on a SemEval2016 English dataset

on restaurants. We achieve in this case great perfor-

mance with a F1-score equal to 80.14% and shows

that our model can be used for ABSA in multiple lan-

guages.

5 CONCLUSION

In this paper we propose a new model named CLC for

ABSA and more precisely for aspect and sentiment

detection.

Several efﬁcient methods have been proposed for

English language in literature. However, only a few

work has been done in French due to the lack of re-

sources. Thus, the main contribution of our work

is that we propose a new model that gives high per-

formance on the French language. CLC is also well

adapted for other language as we proved it for En-

glish.

The strength of the proposed model is its ability to

jointly detect the aspects and the associated sentiment

at the same time and thus fully exploit the correlation

between them.

For the future works, it could be interesting to add

an attention mechanism in CLC which may improve

its performance. There is also a great need to develop

unsupervised models for ABSA to get rid off the de-

pendency on annotations. This would open new per-

spectives to explore since there is a huge amount of

unannotated data available on the web (e.g. customer

reviews on TripAdvisor).

REFERENCES

Apidianaki, M., Tannier, X., and Richart, C. (2016).

Datasets for aspect-based sentiment analysis in

French. In Proceedings of the Tenth International

Conference on Language Resources and Evaluation

(LREC’16).

Brun, C. and Nikoulina, V. (2018). Aspect based sen-

timent analysis into the wild. In Proceedings of

the 9th Workshop on Computational Approaches to

Subjectivity, Sentiment and Social Media Analysis,

WASSA@EMNLP 2018, Brussels, Belgium, October

31, 2018.

Brun, C., Perez, J., and Roux, C. (2016). XRCE at

semeval-2016 task 5: Feedbacked ensemble mod-

eling on syntactico-semantic knowledge for aspect

based sentiment analysis. In Proceedings of the

10th International Workshop on Semantic Evaluation,

SemEval@NAACL-HLT 2016, San Diego, CA, USA,

June 16-17, 2016.

Chen, T., Xu, R., He, Y., and Wang, X. (2017). Improving

sentiment analysis via sentence type classiﬁcation us-

ing bilstm-crf and CNN. Expert Syst. Appl., 72:221–

230.

Chiu, J. P. C. and Nichols, E. (2016). Named entity recogni-

tion with bidirectional lstm-cnns. Trans. Assoc. Com-

put. Linguistics, 4:357–370.

Dong, L., Wei, F., Tan, C., Tang, D., Zhou, M., and Xu, K.

(2014). Adaptive recursive neural network for target-

dependent twitter sentiment classiﬁcation. In Pro-

ceedings of the 52nd Annual Meeting of the Associ-

ation for Computational Linguistics, ACL 2014, June

22-27, 2014, Baltimore, MD, USA, Volume 2: Short

Papers.

Hamdan, H., Bellot, P., and B

echet, F. (2015). Lsislif:

CRF and logistic regression for opinion target extrac-

tion and sentiment polarity analysis. In Proceedings

of the 9th International Workshop on Semantic Eval-

uation, SemEval@NAACL-HLT 2015, Denver, Col-

orado, USA, June 4-5, 2015.

NLPinAI 2021 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

504

Kim, Y. (2014). Convolutional neural networks for sentence

classiﬁcation. CoRR, abs/1408.5882.

Kiritchenko, S., Zhu, X., Cherry, C., and Mohammad, S.

(2014). Nrc-canada-2014: Detecting aspects and sen-

timent in customer reviews. In Proceedings of the 8th

International Workshop on Semantic Evaluation, Se-

mEval@COLING 2014, Dublin, Ireland, August 23-

24, 2014.

Kooli, N. and Pigneul, E. (2018). Analyse de sentiments

base d’aspects par combinaison de r

eseaux profonds

: application

a des avis en franc¸ais (A combination

of deep learning methods for aspect-based sentiment

analysis : application to french reviews). In Actes de la

Conf

erence TALN. CORIA-TALN-RJC 2018 - Volume

1, Rennes, France, May 14-18, 2018.

Kumar, A., Kohail, S., Kumar, A., Ekbal, A., and Bie-

mann, C. (2016). IIT-TUDA at semeval-2016 task

5: Beyond sentiment lexicon: Combining domain de-

pendency and distributional semantics features for as-

pect based sentiment analysis. In Proceedings of the

10th International Workshop on Semantic Evaluation,

SemEval@NAACL-HLT 2016, San Diego, CA, USA,

June 16-17, 2016.

Lin, P. and Luo, X. (2020). A survey of sentiment analysis

based on machine learning. In Zhu, X., Zhang, M.,

Hong, Y., and He, R., editors, Natural Language Pro-

cessing and Chinese Computing - 9th CCF Interna-

tional Conference, NLPCC 2020, Zhengzhou, China,

October 14-18, 2020, Proceedings, Part I.

Ma, D., Li, S., Zhang, X., and Wang, H. (2017). Interactive

attention networks for aspect-level sentiment classiﬁ-

cation. In Proceedings of the Twenty-Sixth Interna-

tional Joint Conference on Artiﬁcial Intelligence, IJ-

CAI 2017, Melbourne, Australia, August 19-25, 2017.

Machacek, J. (2016). Butknot at semeval-2016 task 5: Su-

pervised machine learning with term substitution ap-

proach in aspect category detection. In Proceedings of

the 10th International Workshop on Semantic Evalu-

ation, SemEval@NAACL-HLT 2016, San Diego, CA,

USA, June 16-17, 2016.

Pontiki, M., Galanis, D., Papageorgiou, H., Androutsopou-

los, I., Manandhar, S., Al-Smadi, M., Al-Ayyoub,

M., Zhao, Y., Qin, B., Clercq, O. D., Hoste, V.,

Apidianaki, M., Tannier, X., Loukachevitch, N. V.,

Kotelnikov, E. V., Bel, N., Zafra, S. M. J., and

Eryigit, G. (2016). Semeval-2016 task 5: Aspect

based sentiment analysis. In Proceedings of the

10th International Workshop on Semantic Evaluation,

SemEval@NAACL-HLT 2016, San Diego, CA, USA,

June 16-17, 2016.

Poria, S., Cambria, E., and Gelbukh, A. F. (2016). Aspect

extraction for opinion mining with a deep convolu-

tional neural network. Knowl. Based Syst., 108:42–

49.

Ruder, S., Ghaffari, P., and Breslin, J. G. (2016). INSIGHT-

1 at semeval-2016 task 5: Deep learning for multilin-

gual aspect-based sentiment analysis. In Proceedings

of the 10th International Workshop on Semantic Eval-

uation, SemEval@NAACL-HLT 2016, San Diego, CA,

USA, June 16-17, 2016.

Song, Y., Wang, J., Jiang, T., Liu, Z., and Rao, Y. (2019).

Attentional encoder network for targeted sentiment

classiﬁcation. CoRR, abs/1902.09314.

Tang, D., Qin, B., Feng, X., and Liu, T. (2016a). Ef-

fective lstms for target-dependent sentiment classiﬁ-

cation. In COLING 2016, 26th International Con-

ference on Computational Linguistics, Proceedings of

the Conference: Technical Papers, December 11-16,

2016, Osaka, Japan.

Tang, D., Qin, B., and Liu, T. (2016b). Aspect level sen-

timent classiﬁcation with deep memory network. In

Proceedings of the 2016 Conference on Empirical

Methods in Natural Language Processing, EMNLP

2016, Austin, Texas, USA, November 1-4, 2016.

Thet, T. T., Na, J., and Khoo, C. S. G. (2010). Aspect-

based sentiment analysis of movie reviews on discus-

sion boards. J. Inf. Sci., 36(6):823–848.

Wang, Y., Huang, M., Zhu, X., and Zhao, L. (2016).

Attention-based LSTM for aspect-level sentiment

classiﬁcation. In Proceedings of the 2016 Conference

on Empirical Methods in Natural Language Process-

ing, EMNLP 2016, Austin, Texas, USA, November 1-4,

2016.

Wu, H., Gu, Y., Sun, S., and Gu, X. (2016). Aspect-

based opinion summarization with convolutional neu-

ral networks. In 2016 International Joint Conference

on Neural Networks, IJCNN 2016, Vancouver, BC,

Canada, July 24-29, 2016, pages 3157–3163. IEEE.

Young, T., Hazarika, D., Poria, S., and Cambria, E. (2018).

Recent trends in deep learning based natural language

processing [review article]. IEEE Comput. Intell.

Mag., 13(3):55–75.

CNN-LSTM-CRF for Aspect-Based Sentiment Analysis: A Joint Method Applied to French Reviews

505