Sentiment Analysis of User Comment Text based on LSTM
FENG LI, CHENXI CUI, YASHI HU, LINGLING WANG*
School of Management Science and Engineering,
Anhui University of Finance and Economics,
Bengbu 233030,
CHINA
*Corresponding Author
Abstract: Taking the user-generated Chinese comment dataset on online platforms as the research object, we
constructed word2vec word vectors using gensim and built a sentiment analysis model based on LSTM using
the TensorFlow deep learning framework. From the perspective of mining user comment data on the platform,
we analyzed the sentiment tendency of user comments, providing data support for hotels to understand
consumers' real sentiment tendencies and improve their own service quality. Through analysis of the validation
dataset results obtained by crawling the website, the accuracy of this LSTM model can reach up to 0.89, but
there is still much room for improvement in the accuracy of sentiment analysis for some datasets. In future
research, this model needs further optimization to obtain a stable and more accurate deep-learning model.
Key-Words: Online; Comment Text; LSTM; Sentiment Analysis.
Received: May 8, 2022. Revised: February 12, 2023. Accepted: March 3, 2023. Published: April 5, 2023.
1 Introduction
In recent years, with the rapid development of
Internet technology and the arrival of the era of
universal 5G, Internet applications and big data
computing are involved in every aspect of daily life.
By the end of 2021, the global Internet user base has
reached 4.9 billion. Compared with 2019 (401
billion), the global Internet user case has grown by
19.5%, with an increase of 800 million new Internet
users. Among them, the growth rate of global
Internet users reached 10.2% in 2020, the highest in
a decade. As a result, the Internet generates around
4 PB of data every day, including about 10 billion
text messages. With the increasingly widespread use
of online shopping, the amount of information data
generated by users of e-commerce platforms is also
increasing, with a considerable portion being
emotional evaluations, opinions, and thoughts.
Extracting user sentiment from such a large and
complex volume of textual information and
identifying users' sentiment tendencies has become a
research area of great interest in the field of natural
language processing. Sentiment analysis, also
known as opinion mining, is the process of
classifying text into positive, negative, and neutral
sentiments, [1]. The main purpose of sentiment
analysis is to classify text into positive, negative,
and neutral sentiments. To date, sentiment analysis
has been extensively explored and breakthroughs
have been made in various research methods.
Common methods for sentiment analysis include
traditional sentiment lexicons and machine learning.
The traditional sentiment lexicon method involves
comparing the emotional tendency and information
intensity of the vocabulary in the pre-constructed
sentiment lexicon with the content of the text and
then classifying the text. Traditional sentiment
lexicon methods can use existing high-quality
lexicons, but the lexicon cannot cover all the
vocabulary in all fields, especially with the impact
of the newly emerged internet language. Machine
learning models text classification by modeling
certain features and learning from manually labeled
data, but this method requires some manual feature
engineering.
The content of user comments contains the
subjective emotions of the user. Sentiment analysis
of user message content usually refers to the mining,
parsing, generalisation, and inference of user-
generated content based on machine learning, from
which user attitudes and opinions are obtained.
Sentiment analysis can be divided into chapter-level
sentiment analysis, sentence-level sentiment
analysis, and aspect-level sentiment analysis, with
chapter-level sentiment analysis and sentence-level
sentiment analysis belonging to coarse-grained
sentiment analysis and aspect-level sentiment
analysis belonging to fine-grained sentiment
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
19
Volume 19, 2023
analysis. Coarse-grained sentiment analysis can only
analyse the sentiment of a whole text or paragraph,
but not the multiple perspectives contained in a
large text. In contrast, fine-grained sentiment
analysis can determine the sentiment of individual
entities more accurately, making aspect-level
sentiment analysis a hot research area in the field of
sentiment analysis. Aspect-Based Sentiment
Analysis (ABSA) is a fundamental task in sentiment
analysis, which aims to identify the aspects present
in a sentence and determine the sentiment polarity
of each aspect, 0.
Through research on relevant domestic and
foreign work, we found that the mainstream text
sentiment analysis methods currently mainly include:
sentiment analysis based on sentiment dictionaries,
sentiment analysis based on machine learning, and
sentiment analysis based on deep learning.
Sentiment analysis based on sentiment dictionaries
and rules first compares the text content with
specific entities in the sentiment dictionary to derive
sentiment values, and then the results are weighted
to derive the sentiment tendency of the text. This
method can produce good sentiment analysis results
if the sentiment lexicon is sufficiently rich. A
machine learning algorithm is a generic term for a
class of algorithms that construct a function with a
large amount of data as input, the output of which
can be classification, prediction, etc. And this
function is equally applicable to new sample data.
Machine learning-based sentiment analysis is a
method of constructing a function model into which
a large amount of textual information is input and
extracting features through machine learning
algorithms. KNN, NB, and SVM are common
algorithms used in machine learning. Deep learning
is a practical application of multilayer neural
networks in learning, which is still essentially in the
realm of machine learning, although it can solve
complex problems that are difficult to solve with
traditional machine learning, so it has been singled
out as a separate discipline. Common models of
deep learning include CNN, RNN, Transformer,
GRU, and LSTM. The underlying structure of a
neural network is shown in Figure 1 and contains an
input layer, a hidden layer, and an output layer.
Each neuron in the input layer can be used as a
feature of an object, the hidden layer may have
multiple layers and it will transform the information
from the input into something that can be used in the
output layer, and the output layer transforms the
results of the hidden layer into the desired result.
Fig. 1: Neural network structure
2 Related Works
2.1 Studies Related to Sentiment Analysis
2.1.1 Sentiment Analysis Methods based on
Sentiment Dictionaries
The earliest English sentiment lexicon that appeared
is SentiWordNet, in addition to the commonly used
ones, such as General Inquirer, Opinion Lexicon,
and MPQA. Chinese emotion dictionaries are
widely used, such as HowNet and Dalian University
of Technology's Chinese emotion vocabulary
ontology database. In addition, 0, proposed a
sentiment lexicon for Vietnamese, which includes
more than 100,000 Vietnamese emotional
vocabularies. The slang sentiment dictionary
SlangSD, which was built by [4], from web
resources, is also effective in identifying the
sentiment of users. [5], used a constructed
dictionary of network words, degree adverbial
dictionaries, negation dictionaries, and other related
dictionaries, and trained them with the help of
Weibo texts to derive updated sentiment values. [6],
proposed a sentiment classifier that trains
incremental words from time-varying distributional
word vectors, automatically extracts continuously
updated sentiment vocabulary from Twitter streams,
and obtains a time-varying sentiment lexicon based
on incremental word vectors. [7], integrated emojis,
modifiers, and domain-specific terms to analyse
comments posted by online communities,
overcoming the limitations of previous methods.
Compared to the general approach, sentiment
analysis was greatly improved by integrating
modifiers, emojis, negation words, and domain-
specific terms. However, the sentiment analysis
method based on sentiment dictionaries relies too
much on sentiment dictionaries, and in today's
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
20
Volume 19, 2023
information age, a single sentiment dictionary
cannot make accurate judgments, while building a
more complete and diverse dictionary can be labor-
intensive.
2.1.2 Machine Learning-based Sentiment
Analysis Methods
[8], compared the results of decision trees, Bernoulli
NB (BNB), Maximum Entropy (ME), support
vector machines (SVM), and multinomial naive
Bayes (MNB) in sentiment classification, and found
that multinomial naive Bayes obtaining the best
results of 88. 5%. [9], constructed a sentiment
analyzer based on SVM and naive Bayes to analyze
Twitter data and compared it with a sentiment
analyzer using only SVM or NB. [10], proposed an
optimized sentiment analysis framework (OSAF),
which uses SVM lattice search techniques and
cross-validation. [11], proposed an emoticon-based
sentiment analysis method and discussed the role of
symbolic expressions in sentiment analysis. [12],
proposed a computational algorithm for semantic
analysis based on the WordNet linguistic English
lexicon training set, using a combination of machine
learning algorithms SVM and NB to automatically
detect strongly associated negative tweets.
Although machine learning-based sentiment
analysis has made progress compared to lexicon-
based sentiment analysis, it still requires manual
labeling of text and subjective factors can affect the
final result. Traditional machine learning requires
high model requirements, and if the model is not
efficient, it is difficult to adapt to the era of
exploding information. In addition, traditional
machine learning has difficulty using contextual
information in sentiment analysis, which also affects
accuracy.
2.1.3 Deep Learning-based Sentiment Analysis
Methods
A sentiment analysis method based on deep learning
can automatically learn deep features from a large
amount of text information, and the sentiment
analysis is effective and the model is highly
adaptable without human intervention during the
learning process. [13], proposed a Restricted
Boltzmann Machine (RBM) based rule model for
sentiment analysis of sentences. [14], proposed a
restricted data framework using RNN as a
framework to train a single model using the largest
dataset of languages and reuse it for languages with
limited datasets. This framework has good results
for sentiment analysis of small languages. LSTM is
a special structure of RNN, and to improve the
training speed and reduce computational cost and
time, [15], proposed an attention-based LSTM
oriented aspect-level sentiment memory network
classification for sentiment classification based on
LSTM, [16], proposed a streamlined LSTM with six
different parameters and compared the performance
differences between these LSTMs using the Twitter
dataset to establish the best set of parameters for the
LSTM, [17], proposed a new sentiment analysis
scheme based on Twitter and Weibo data, focusing
on the impact of expressions on sentiment, and
training an emotion classifier by attending to these
binary expressions, embedded in an attention-based
long- and short-term memory network, which is a
good guide for sentiment analysis. Because of the
lower human input as well as the higher accuracy,
deep learning-based sentiment analysis methods
have become a hot research topic in recent years.
2.1.4 Analysis of Irony
It is easy to find that there are a lot of phenomena of
irony and sarcasm on online platforms, and the
emotion implied by such statements is often the
opposite of the surface meaning of the statements.
Therefore, the analysis of ironic statements and the
analysis of the deeper meaning of the statements
will help to determine the emotional polarity of the
text. [18], achieved good results in experiments with
four machine learning methods by improving the
sentiment analysis process and decision-making
process and crawling data on Twitter, linear SVC
(accuracy=83%, f1-score=0.81), logistic regression
(accuracy=83%, f1-score=0.81), Naïve Bayes
(accuracy=74%, f1- score=0.73) and random forest
classifier (accuracy=80%, f1-score=0.81). Some
authors, [19], found that previous research on
sarcasm detection has mostly been conducted using
natural language processing techniques, without
considering the context, user's expression habits, etc.
Therefore, a two-channel convolutional neural
network was used to analyze the semantics of the
target text, as well as its emotional context, and to
extract the user's expression habits using an
attention mechanism. The effectiveness of the
method is confirmed by experiments on several
datasets, and it can effectively improve the
performance of the irony detection task.
2.1.5 Implicit Sentiment Analysis
Implicit sentiment analysis is a special part of the
sentiment analysis field because of the lack of
sentiment vocabulary and the ambiguity of
sentiment polarity, which is a difficult area of
research at this stage. Combing the literature on
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
21
Volume 19, 2023
implicit sentiment analysis at this stage, it is found
that the current research is very limited. [20], found
that previous Graph Convolutional Networks
(GCNs) used for the study of sentiment analysis
problems had difficulty in effectively using
contextual context or often ignored the
dependencies between phrases. Therefore, they
proposed a context-specific heterogeneous graph
convolutional network (CsHGCN) based on this,
and experimental results showed that the model
could effectively identify target emotions in
sentences.
2.1.6 Aspect-level Sentiment Analysis
Aspect-Based Sentiment Analysis (ABSA), an
actively challenging part of the sentiment analysis
field, aims to identify and analyze the fine-grained
sentiment polarities towards particular aspects.
[21], proposed a new neural network-based
framework to analyze the sentiment of aspect targets
in comments. This framework captures distant
textual sentiment information through a multi-
attentive mechanism, employing a non-linear
combination with recurrent neural networks to
enhance the expressive power of the model,
allowing it to handle more complex semantic
problems. The performance of this model is also
validated on four datasets (two from SemEval2014
(restaurant and laptop reviews), a Chinese news
review dataset, and Twitter datasets).
[22], found that most previous prediction
methods used long- and short-term memory and
attention mechanisms to analyze the emotional
polarity of the target of interest, and that such
methods tended to be more complex and required
more training time. Therefore, it was proposed to
group the previous methods into two subtasks:
aspect-category sentiment analysis (ACSA) and
aspect-item sentiment analysis (ATSA). A model
based on gating mechanisms and convolutional
neural networks is also proposed, which is more
accurate and effective. The method firstly uses a
new gating unit, Tanh-ReLU, to selectively output
sentiment features based on a given entity or aspect;
this architecture is simpler than the attention layer
used in existing models; secondly, the computations
of this model are easily deserialized during training
and the gating unit works independently, and finally,
experiments on the SemEval dataset validate the
effectiveness of the model.
Arabic poses several challenges for the task of
sentiment analysis in Arabic because of its complex
grammatical structure and the lack of relevant
resources. Some scholars have taken the sentiment
analysis of aspects of Arabic as a research direction,
[23], and used a composite model combining a long
short-term memory (LSTM) model and a
convolutional neural network (CNN) to analyze the
sentiment of Arabic tweets. For the Arabic
sentiment tweet dataset (ASTD), this model scored
64.46% on F1, outperforming other deep learning
models; some scholars, [24], research using two
different long short-term memory (LSTM) neural
networks for aspect-level sentiment analysis of
Arabic hotel reviews. The first is an aspect-OTEs
oriented LSTM for aspect sentiment polarity
classification as sentiment polarity markers, and the
second is a character-level bidirectional LSTM
along with a conditional random field classifier (Bi-
LSTM-CRF) for aspect opinion target expression
(ballot) extraction. This method was evaluated using
a reference dataset of Arabian hotel reviews and the
results showed that this method outperformed the
baseline study on both tasks by 6% and 39%
respectively.
2.2 Relevant Research Techniques
Sentiment analysis of text content is the complete
process of text preprocessing such as word
segmentation, stop-word removal, and named entity
recognition on the target text, followed by text
vectorization, feature engineering, model training,
classifier, and other processes to derive sentiment
tendency labels. A flowchart of text classification is
presented in Figure 2.
Fig. 2: Flowchart of text classification
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
22
Volume 19, 2023
2.2.1 Text Pre-Processing
The implementation of text preprocessing mainly
involves word segmentation of txt format texts. This
experiment uses the jieba Chinese word splitting
tool, a widely used and effective Chinese word
splitter with exact mode, full mode, and search
engine mode that allows for precise segmentation of
text sentences, fast scanning of the entire content,
and secondary recall segmentation of long words. It
is also possible to improve the segmentation effect
by manually defining proper nouns in the text. In
practice, there are a large number of intonational
auxiliaries, personal pronouns, and other words that
are not related to emotional tendencies, which can
be filtered by building a deactivation dictionary. In
addition, there are different dictionaries for different
domains, [25]. The dictionaries are designed to
include new words and specialized words that are
unique to the field.
2.2.2 Text Vectorization
Text vectorization refers to converting Chinese text
content that cannot be recognized by a computer
into a vector form with digitized features that can be
recognized by a computer. In this paper, word2vec
is used to complete text vectorization, transforming
the text preprocessed dataset into a vector with
uniform dimensions, thus completing the
simplification of shifting data from high latitude to
low latitude. Depending on the definition of output
and input, two algorithms can be classified: Skip-
gram and CBOW. The CBOW algorithm is a three-
layer neural network that predicts target words from
contextual words, defining the words of the context
in which a word is located as input and itself as
output, using a corpus of corpora for training, and
calculating the vector values of the context in the
projection layer and summing them to output
information about the target word. The Skip-gram
algorithm reverses the causality of CBOW by
defining the words in the context of the target word
as the output, and the words themselves as the input,
predicting the information of the contextual words
with the help of the target word.
2.2.3 Recurrent Neural Network (RNN)
Recurrent Neural Networks are a class of sequences
data as input and perform in the direction of
evolution of the sequence recursively and all nodes
(recurrent units) are connected in a chain-like
manner in recurrent neural networks. Recurrent
Neural Networks have memory, share parameters,
and are Turing-complete and are therefore very
useful in the analysis of sequential non-linear
features. Recurrent neural networks are used in
natural language processing such as speech
recognition, language modelling, and machine
translation, and also for various time-series
predictions. The introduction of convolutional
neural networks constructed recurrent neural
networks that can process sequential inputs
containing computer vision problems.
2.2.4 Long and Short-Term Memory
Long short-term memory (LSTM) is a special kind
of RNN that was designed to address the problem of
gradient vanishing and explosion during the training
of long sequences. LSTM is a variant of RNN, with
the core concept of cell states and "gate" structure.
Cell states are the equivalent of information
transmission paths that allow information to be
passed along in a sequence. You can think of it as
the 'memory' of the network. Theoretically, the cell
states can pass on information relevant to the
sequence processing all the way through. Thus, even
information from earlier time steps can be carried to
cells at later time steps, which overcomes the effects
of short-term memory. Information is added and
removed by means of 'gate' structures, which are
trained to learn which information to keep or forget.
In recent years, recurrent neural networks have been
used in speech recognition, image processing, ECG
arrhythmia classification, and natural language
processing, and so on [26]. The LSTM model is
shown in Figure 3.
Fig. 3: LSTM structure
3 Algorithm
Due to the over-reliance on sentiment lexicons,
simple lexicons cannot accurately discriminate
sentiments, while constructing a more complex and
diverse lexicon will require a large amount of
manpower and resources. Although sentiment
analysis based on machine learning has made
progress compared to lexicon-based sentiment
analysis, subjective factors have a significant impact
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
23
Volume 19, 2023
during manual calibration. Moreover, traditional
machine learning models require high model
accuracy, and the explosion of information in
today's world makes it difficult for models to adapt
perfectly to complex and varied needs. In addition,
traditional machine learning has difficulty utilizing
contextual information, which can affect accuracy in
sentiment analysis. Deep learning-based sentiment
analysis methods can automatically learn deep
features from a large amount of text information,
with good sentiment analysis effects and strong
model adaptability, without the need for human
intervention in the learning process. Due to the low
efficiency and quality of traditional methods, people
have begun to use deep learning to construct
network models for text classification tasks. [30],
reviewed more than 150 deep learning-based text
classification models developed in recent years in
their review and discussed their technical
contributions, similarities, and advantages.
Therefore, this paper chose a deep learning-based
sentiment analysis method to complete the
sentiment judgment of text information.
Common deep-learning models include CNN,
RNN, Transformer, GRU, and LSTM. Traditional
CNN models may not activate neurons that
recognize the same object slightly differently due to
translational invariance, i.e., changes in the
orientation or position of the same object. Moreover,
the pooling layer causes a significant loss of
valuable information, ignoring the correlation
between local and global features. Therefore, CNN
models are difficult to accurately judge the precise
textual sentiment. Although RNN models can
consider historical information during calculation
and share weights over time compared to CNN
models, their computation speed is slow and cannot
consider any future input of the current state. In
addition, RNN models often suffer from gradient
disappearance and explosion because it is difficult
to capture long-term dependencies, and
multiplication gradients can decrease or increase
exponentially with the number of layers. Although
GRU models can effectively alleviate the problem
of gradient explosion in RNN models, compared to
GRU models, LSTM models have more parameters,
stronger functionality, and stronger expressive
power.
LSTM has a similar working mechanism to
RNN, but its implementation of more refined
internal processing units enables effective storage
and updating of contextual information. Due to its
excellent properties, LSTM has been used in many
tasks related to sequence learning, such as speech
recognition, [31], language models, [32], part-of-
speech tagging, [35], and machine translation, [36].
Therefore, considering all factors, this paper uses
LSTM as the deep learning model for sentiment
analysis.
3.1 Recurrent Network Model
RNNs, or Recurrent Neural Networks, excel in
processing sequences of data where context is
essential. One of the distinguishing features of
RNNs is their ability to create directed loops
between nodes, [38]. Examples of sequence data
that RNNs can handle well include speech
recognition, language prediction, garbage image
classification, [39], and stock data analysis, [40].
Since the data at each node in the sequence is
related to the preceding and subsequent data points,
RNNs can capture these dynamic relationships. By
retaining previous information and using it as input
for subsequent nodes, RNNs are ideal for analyzing
time-sequenced data.
3.2 RNN Model Gradient Disappearance
Phenomenon
[41], proposed that standard RNNs suffer from
gradient vanishing, which refers to the vanishing of
gradients in RNNs for more distant time steps. The
BPTT method is used for backpropagation in RNNs,
where the gradient of loss against parameter W is
equal to the sum of the derivatives of loss against W
at each time step. This can be expressed
mathematically as a formula.:
ℎ
ℎ
ℎ
ℎℎ
ℎ
The calculation in the above equation is more
complex and is based on a continuous derivative of
the complex function.
ℎ
ℎ ℎ
ℎ
ℎ
ℎ is the partial derivative of the current
hidden state with respect to the previous hidden
state.
ℎ
ℎ ℎ
Suppose that a time step j is (t-j) moments away
from time step t. So:
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
24
Volume 19, 2023
ℎ
ℎ ℎ
If t-j is large, that is, j is far from the t time step,
when>1, a gradient explosion problem arises
and <1, there is a gradient disappearance
problem. And when t-j is small, there is no gradient
disappearance/gradient explosion problem. In
summary, the gradient of j farther away from time
step t will vanish and j does not affect the final
output has no effect on the final output. This
means that there can be no long-term dependence on
RNN.
3.3 Gradient Disappearance Phenomenon
To address the problem of long-term dependencies,
[42], proposed a Long Short-Term Memory (LSTM)
network, which performs much better than RNN,
especially in long-distance dependent tasks, [43].
The LSTM was originally designed so that the bias
derivative of the current memory unit with respect
to the previous memory unit would be constant. As
in the original version of the LSTM in 1997, the
memory cell update formula was
Later, to avoid the wireless growth of memory
cells, Σφάλμα! Το αρχείο προέλευσης της
αναφοράς δεν βρέθηκε., later refined the LSTM
cell by introducing the "forget gate". The updated
formula is:
The value of the partial derivatives at this
moment is:
Although is a value in the interval [0,1], not in
the sense of satisfying the bias of the current
memory cell to the previous memory cell as a
constant. However, it is common to set a large bias
term to the forgetting gate such that the forgetting
gate is closed in most cases and open only in a few
cases. Recall the formula for the forgetting gate,
here we have added the bias b.
ℎ
The forget gate is closed when it tends to 1 and
opened when it tends to 0. By setting a large bias
term, most forget gates tend to 1. By setting a large
bias term, most of the forgetting gates converge to 1.
This also alleviates the problem of gradient
disappearance due to fractional multiplication.
4 Sentiment Analysis based on LSTM
4.1 Data and Processing
In this paper, the dataset is based on the comment
corpus collated by Tan Songbo, with 2000 positive
and negative examples each, which is a relatively
small dataset. Examples are shown in Table 1.
Moreover, Table 2 presents model parameter
settings.
Table 1. Example of ChnSentiCorp data
Positive
Negative
It is a very nice 5-star
hotel, the rooms are large,
the facilities are new, and
the location is convenient to
the financial center, so I
would consider staying
there again.
Depressed!!!
Angry!!! I don't
understand that the fiber
optic is even slower than
the internet speed in
Shanghai Jinjiang Star,
don't go to this place if
you want fast internet
speed at night!!!!
The room was clean,
the facilities were ok, the
furniture was a bit old. The
business room has a good
floor front desk and the
price point is relatively low
for a 4-star.
The room was never
arranged to have a frontal
lake view, especially as
the standard of the
reception was really poor,
with grumbling and
expressionless faces.
The hotel was clean,
the waiter would
recommend me to the ladies'
non-smoking floor, the
facilities were better, and
the dim sum in the
restaurant tasted ok.
Not as bad as a good
2-star or no-star hotel
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
25
Volume 19, 2023
Table 2. Model parameter settings
Dim
Words
Buffer
LSTM_size
Dropout
Epochs
Batch size
300
300
3500
32
0.5
25
20
Word vectors: This experiment uses open-
source word vectors and Chinese-word-vectors The
Word Vector is a Word Vector trained from the
Zhihu corpus.
In this work, the data was divided into a training
set and a test set in a ratio of 4:1. For the training
and validation sets, the following format was
followed when producing the training data: In the
text file, each row is the input for one sample, where
each paragraph is commented on for one line and
separated from the word by space using jieba.
4.2 Measurement Criteria
In this paper, recall, accuracy, precision, and F1
values are used as experimental measures and
positive texts are used to refer to texts with positive
affective tendencies and negative texts to refer to
texts with negative affective tendencies. In the
above confusion matrix, TP is the number of texts
correctly classified as positive; FN is the number of
texts incorrectly classified as positive; FP is the
number of texts incorrectly classified as negative;
and TN is the number of texts correctly classified as
negative.
Precision is the percentage of texts judged to be
of a certain type that is correctly judged.
The recall is the percentage of texts that are
actually of a certain type that are judged to be
correct.
The F1 value is the summed mean value of
precision and recall, which corresponds to the
combined precision and recall evaluation metric.
Accuracy is the percentage of correctly judged
texts out of all texts.
4.3 Parameter Settings
The experiment in this paper used the open-source
word embedding model from Zhihu to train text
information into 300-dimensional word vectors. The
parameters of the LSTM model were set as follows:
the maximum word count was set to 300 (setting the
dimension too high would result in longer training
time); a buffer zone of 3500 was reserved; the
regularization parameter was set to 0.5; the batch
size was set to 20; and the algorithm worked 25
times on the entire training dataset.
4.4 Results
Table 3. Experiment results using LSTM
Textual
Emotional
Tendencies
Results
Positive
Negative
support
Positive texts
TP: 865
FN: 135
1000
Negative text
FP: 91
TN: 909
1000
Table 4. LSTM model processing data results
precision
recall
f1-
score
support
POS
0.90
0.87
0.88
1000
NEG
0.87
0.91
0.89
1000
micro avg
0.89
0.89
0.89
2000
macro avg
0.89
0.89
0.89
2000
weighted
avg
0.89
0.89
0.89
2000
4.4.1 Model Training Results
In this paper, 5000 positive and 5000 negative
emotion texts were used for the training of the
model, which were divided into a training set and a
test set according to 4:1, with 8000 texts in the
training set, 4000 positive and 4000 negative
emotions texts in the training set, and 2000 texts in
the test set, containing positive and the test set
contains 2000 texts, including 1000 positive and
1000 negative texts. After processing the LSTM
model, the following results were obtained.
In Table 3 we present the experiment results
using LSTM. Similarly, in Table 4 we present the
LSTM model processing the data results of our
paper. Specifically, regarding Table 4 properties, we
specify the following:
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
26
Volume 19, 2023
1. Macro average macro avg: sums the
accuracy, recall, and F1 values for each
category to find the average.
2. Micro avg builds a global confusion
matrix for each instance in the dataset,
regardless of category, and then
calculates the corresponding metric.
3. weighted avg: an improvement on
macro-averaging, considering the
number of samples in each category as a
proportion of the total sample.
4.4.2 Validation of the Dataset Results
By importing a corpus of e-commerce reviews from
Baidu's library into the trained model, containing
1000 positive and negative texts each, the 2000 texts
were divided into 20 groups of data, and the
accuracy, recall, F1 value, and accuracy of these 20
groups were calculated. The following graphs were
generated from the results.
According to the analysis of the above graphs,
we can find that: the accuracy of the positive text
can reach a maximum of 0.98 and a minimum value
of 0.82; the accuracy of the negative text can reach a
maximum of 0.92 and a minimum value of 0.44; the
recall of the positive text can reach a maximum of
0.91 and a minimum value of 0.64; the recall of the
negative text can reach a maximum of 0.98 and a
minimum value of 0.68; the accuracy of the positive
text The maximum F1 value for positive text is
0.8727 and the minimum value is 0.7189; the
maximum F1 value for negative text is 0.8383 and
the minimum value is 0.5626; the accuracy of this
LSTM model can reach up to 0.89.
A comprehensive analysis of this LSTM model
leads to the conclusion that the accuracy of this
LSTM model still needs to be improved and further
improvements are needed to achieve more accurate
sentiment propensity analysis.
5 Conclusion
In this paper, the sentiment tendency analysis of e-
commerce platform reviews is carried out by the
LSTM model, which is trained and validated by an
open dataset downloaded from the web. Our
research findings are summarized in Figure 4
regarding our experimental results for the LSTM
model validation dataset. Moreover, Figure 5
showcases the accuracy, recall, and F1 values for
forward text whereas Figure 6 is for negative text.
Lastly, Figure 7 presents the overall accuracy of the
studied data sets.
The results show that the classification accuracy
of this LSTM model can reach a maximum of 0.89,
but there is still much room for improvement. The
LSTM model implemented in this paper aims to
judge the sentiment tendency of user-generated
reviews on e-commerce platforms, to perform
sentiment analysis on reviews on e-commerce
platforms, and to provide a proven method for e-
commerce platforms to judge the sentiment polarity
of user reviews and extract keywords in the process
of investigating user feedback, to provide data
support for merchants to understand consumers'
needs and real reviews, and to improve service
quality in a targeted manner. It provides data
support. Sentiment analysis of user reviews can
effectively find out whether users identify with a
shop, observe how much they like the product, help
the management of the e-commerce platform to
discover the strengths and weaknesses of the shop,
improve the level of service and enhance user
satisfaction.
The collective amount of data taken in this
experiment is not large enough for effective analysis
of non-semantic symbols and expressions, the
model training takes too long, and there are
individually large differences in the process of
analyzing the accuracy of the validation set. The
analysis of emoji information, the use of multiple
parameters, and the optimization of the model will
be the next research directions in the future. In
subsequent research, a comparison between the
optimized LSTM model and other neural network
deep learning models will also be obtained with the
increasing capability of text information recognition
and generalization.
Based on the content of this paper, future
research can be conducted in four areas. First, it can
further optimize the sentiment analysis model and
try to use more efficient and accurate deep learning
models, such as pre-trained language models such as
BERT and GPT, and combine with attention
mechanisms to improve the model's performance.
Second, it is necessary to explore how to deal with
the challenges of semantic complexity and
ambiguity in Chinese sentiment analysis, further
improving the accuracy and robustness of the model.
Finally, it is necessary to consider the evolution of
emotions and contextual factors to more accurately
determine the user's emotional tendencies. These
research directions will help further improve the
effectiveness of text sentiment analysis based on
LSTM and make it more applicable to practical
scenarios.
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
27
Volume 19, 2023
Fig. 4: Summary of experimental results for the LSTM model validation dataset
Fig. 5: Accuracy, recall, F1 values for forward text
0
0,2
0,4
0,6
0,8
1
1,2
8 4 9 10 12 14 518 20 11 17 1 6 7 13 316 215 19
Summary of results for the validation dataset
P(POS) R(POS) F1(POS) P(NEG)
R(NEG) F1(NEG) Accuracy
0
0,2
0,4
0,6
0,8
1
1,2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Positive text accuracy, recall, F1 values
P(POS) R(POS) F1(POS)
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
28
Volume 19, 2023
Fig. 6: Accuracy, recall, F1 values for negative text
Fig. 7: Accuracy of data sets
0
0,2
0,4
0,6
0,8
1
1,2
12345678910 11 12 13 14 15 16 17 18 19 20
Accuracy, recall, F1 values for negative text
P(NEG) R(NEG) F1(NEG)
0,71
0,71
0,72
0,73
0,73
0,74
0,74
0,75
0,75
0,76
0,80
0,83
0,83
0,84
0,85
0,86
0,87
0,88
0,88
0,89
0 0,2 0,4 0,6 0,8 1
2
16
20
17
11
5
15
14
12
8
Accuracy
Accuracy
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
29
Volume 19, 2023
Acknowledgment:
We thank the anonymous reviewers and editors for
their very constructive comments. This work was
supported in part by the Undergraduate teaching
quality and teaching reform project of Anhui
University of Finance and Economics under Grant
No. acszjyyb2021035.
References:
[1] Chaturvedi I, Ragusa E, Gastaldo P, et al. Bayesian
network based extreme learning machine for
subjectivity detection, Journal of the Franklin
Institute, 2017, 355(4): 1780-1797.
[2] Zhou J, Huang J X, Chen Q, et al. Deep learning for
aspect-level sentiment classification: survey, vision,
and challenges, IEEE Access, 2019,7: 78454-78483.
[3] Thien K T, Thi P T. A hybrid approach for building
a Vietnamese sentiment dictionary. Journal of
Intelligent & Fuzzy Systems,2018,35(1):1-12.
[4] Wu L, Morestatter F, Liu H, Et al. SlangSD:
building, expanding and using a sentiment
dictionary of slang words for short text sentiment
classification . Language Resources and
Evaluation,2018,52:839-852.
[5] Zhang S X, Wei Z L, Wang Y, et al. Sentiment
analysis of Chinese micro-blog text based on
extended sentiment dictionary . Future Generation
Computer Systems,2018,81:395-403.
[6] Bravo-Marques F, Khanchandani A, Pfahringer B.
Incremental word vectors for time-evolving
sentiment lexicon induction. Cognitive
Computation,2021,14:425-441.
[7] Asghar M Z, Khan A, Ahmad S, et al. Lexicon-
enhanced sentiment analysis framework using rule-
based classification scheme. PLoS One,2017,12(2):
e0171649.
[8] Baid P, Gupta A, Chaplot N. Sentiment analysis of
movie reviews using machine learning techniques.
International Journal of Computer
Applications,2017,179(7):45-49.
[9] Hasan A, Moin S, Karim A, et al. Machine learning-
based sentiment analysis for twitter accounts.
Mathematical and Computational
Applications,2018,23(1):11.
[10] Ahmad M, Aftab S, Bashir S, et al. SVM
optimization for sentiment analysis. International
Journal Advanced Computer Science and
Applications,2018,9(4):393-398.
[11] Mathapati S, Nafeesa A, Manjula S H, et al.
OTAWE Optimized topic-adaptive word expansion
for cross domain sentiment classification on tweets.
Advances in Machine Learning and Data Science.
Singapore: Springer,2018,705:213-224.
[12] Birjali M, Benihssane A, Erritali M. Machine
learning and semantic sentiment analysis based
algorithms for suicide sentiment prediction in social
networks. Procedia Computer Science,2017,113:65-
72.
[13] Dwivedi R K, Aggarwal M, Keshari S K, et al.
Sentiment analysis and feature extraction using rule-
based model (RBM). Proceedings of the 2019
International Conference on Innovative Computing
and Communications. Cham: Springer,2019:57-63.
[14] Can E F, Ezencan A, Can F. Multilingual sentiment
analysis: an RNN-based framework for limited data
[EB/OL]. [2018-06-08]. https://arxiv. org/pdf/1806.
04511. pdf.
[15] Wang Y Q, Huang M L, Zhu X Y, et al. Attention-
based LSTM for aspect-level sentiment
classification. Proceedings of the 2016 Conference
on Empirical Methods in Natural Language
Processing. Stroudsburg, PA: Association for
Computational Linguistics,2016:606-615.
[16] Gopalakrishnan K, Salem F M. Sentiment analysis
using simplified long short-term memory recurrent
neural networks. https://arxiv. org/abs/2005.
03993v1.
[17] Chen Y X, Yuan J B, You Q Z, et al. Twitter
sentiment analysis via bi-sense emoji embedding
and attention-based LSTM . Proceedings of the 26th
ACM international conference on Multimedia. New
York:ACM,2018:117-125.
[18] Chaudhary S, Kakkar M. Sarcasm detection
technique on twitter data with natural language
processing.Proceedings of International Conference
on Big Data, Machine Learning and their
Applications. Singapore: Springer,2021:283-293.
[19] Du Y, Li T, Pathan M S, et al. An effective sarcasm
detection approach based on sentimental context and
individual expression habits. Cognitive
Computation,2021(1):1-13.
[20] Zuo E, Zhao H, Chen B, et al. Context-specific
heterogeneous graph convolutional network for
implicit sentiment analysis. IEEE
Access,2020,8:37967-37975.
[21] Chen Peng, Sun Zhongqian. Recurrent attention
network on memory for aspect sentiment
analysis.Proceedings of the 2017 Conference on
Empirical Methods in Natural Language
Processing,2017.
[22] Wei X, Tao L. Aspect based sentiment analysis with
gated convolutional networks.
arXiv:1805.07043,2018.
[23] Heikal M, Torki M, Elmakky N. Sentiment analysis
of arabic tweets using deep learning. Procedia
Computer Science,2018,142:114-122.
[24] Al-Smadi M, Talafha B, Al-Ayyoub M, et al. Using
long short-term memory deep neural networks for
aspect-based sentiment analysis of Arabic reviews.
International Journal of Machine Learning and
Cybernetics,2018,10:2163-2175.
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
30
Volume 19, 2023
[25] Cai Y, Yang K, Huang D P, et al. A hybrid model
for opinion mining based on domain sentiment
dictionary. International Journal of Machine
Learning and Cybernetics,2019,10:2131-2142.
[26] Graves A, Mohamed A, Hinton G. Speech
Recognition with Deep Recurrent Neural
Networks.Proceedings of International Conference
on Acoustics, Speech and Signal Processing, 2013:
6645-6649.
[27] B. Hou, J. Yang, P. Wang, and R. Yan, “LSTM
Based Auto-Encoder Model for ECG Arrhythmias
Classification”, IEEE Transactions on
Instrumentation and Measurement, 2020, pp. 1232-
1240.
[28] Gregor K, Danihelka I, Graves A, Rezende D J,
Wierstral D. DRAW: A Recurrent Neural Network
for Image Generation.Proceedings of International
Conference on Machine Learning, 2015: 1462-1471.
[29] Mikolov T, Kombrink S, Deoras A, Burget L,
Cernocky A J H. RNNLM-Recurrent Neural
Network Language Modeling Toolkit.Proceedings
of Automatic Speech Recognition and
Understanding Workshop, 2011: 196-201.
[30] Minaee S, Kalchbrenner N, Cambria E, et al. Deep
learning based text classification: a comprehensive
review. Computation and Language, 2020, 8(5):
85616- 85638.
[31] Graves A, Jaitly N. Towards end-to-end speech
recognition with recurrent neural networks.
Proceedings of the 31st International Conference on
Machine Learning. Beijing: JMLR, 2014: 1764 -
1772.
[32] Akase S, Suzuki J, Nagata M. Input-to-output gate
to improve RNN language models. arXiv Preprint,
2017, 2017: arXiv: 1709. 08907.
[33] Miyamoto Y, Cho K. Gated word-character
recurrent language model. arXiv Preprint, 2016,
2016: arXiv: 1606. 01700.
[34] Jozefowicz R, Vinyals O, Schuster M, et al.
Exploring the limits of language modeling. arXiv
Preprint, 2016, 2016: arXiv: 1602. 02410.
[35] CHO K, Van Merrienboer B, Gulcehre C, et al.
Learning phrase representations using RNN
encoder-decoder for statistical machine translation.
arXiv Preprint, 2014, 2014: arXiv: 1406. 1078.
[36] Bahdanau D, Cho K, Bengio Y. Neural machine
translation by jointly learning to align and translate.
arXiv Preprint, 2014, 2014: arXiv: 1409. 0473.
[37] Wu Y, Schuster M, Chen Z, et al. Google's neural
machine translation system: Bridging the gap
between human and machine translation. arXiv
Preprint, 2016, 2016: arXiv: 1609. 08144.
[38] Mou, Lichao, Pedram Ghamisi, and Xiao Xiang
Zhu. "Deep recurrent neural networks for
hyperspectral image classification." IEEE
Transactions on Geoscience and Remote Sensing
55.7 (2017): 3639-3655.
[39] Feng Li, Lingling Wang, "Application of Deep
Learning Based on Garbage Image Classification,"
WSEAS Transactions on Computers, vol. 21, pp.
277-282, 2022
[40] Feng Li, Lingling Wang, "Case-Based Teaching for
Stock Prediction System Based on Deep Learning,"
WSEAS Transactions on Business and Economics,
vol. 19, pp. 1325-1331, 2022
[41] Bengio Y, Simard P, Frasconi P. Learning long-term
dependencies with gradient descent is difficult.
IEEE Transactions on Neural Networks, 1994, 5(2):
157 - 166.
[42] Hochreiter S, Schmidhuber J. Long short-term
memory. Neural Computation, 1997, 9(8): 1735 -
1780.
[43] Gers F A, Schraudolph N N. Learning precise timing
with LSTM recurrent networks . Journal of Machine
Learning Research, 2002,3(1): 115 - 143.
[44] Gers F A, Schmidhube R J, Cummins F. Learning to
forget: continual prediction with LSTM . Neural
Computation, 2000, 12(10): 2451 - 2471.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
-Feng Li, Chenxi Cui carried out the simulation and
the optimization.
-Yashi Hu, Lingling Wang have organized and
executed the experiments of Section 4.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This work was supported in part by the
Undergraduate teaching quality and teaching reform
project of Anhui University of Finance and
Economics under Grant No. acszjyyb2021035.
Conflict of Interest
The authors declare that the research was conducted
in the absence of any commercial or financial
relationships that could be construed as a potential
conflict of interest.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.3
Feng Li, Chenxi Cui, Yashi Hu, Lingling Wang
E-ISSN: 2224-3488
31
Volume 19, 2023
