A METHOD OF FINE-GRAINED SHORT

TEXT SENTIMENT ANALYSIS BASED ON

MACHINE LEARNING
G. Chang∗, H. Huo†

Abstract: Text sentiment analysis plays an important role in social network in-
formation mining. It is also the theoretical foundation and basis of personalized
recommendation, circle of interest classification and public opinion analysis. In
view of the existing algorithms for feature extraction and weight calculation, we
find that they fail to fully take into account the influence of sentiment words. There-
fore, this paper proposed a fine-grained short text sentiment analysis method based
on machine learning. To improve the calculation method of feature selection and
weighting and proposed a more suitable sentiment analysis algorithm for features
extraction named N-CHI and weight calculation named W-TF-IDF, increasing the
proportion and weight of sentiment words in the feature words Through experimen-
tal analysis and comparison, the classification accuracy of this method is obviously
improved compared with other methods.

Key words: fine-grained sentiment analysis, sentiment information classification,

feature extraction, weight calculation
Received: November 28, 2017 DOI: 10.14311/NNW.2018.28.019
Revised and accepted: June 20, 2018

1. Introduction
With the rapid development of the Internet, the function of the network is more and
more comprehensive, and the use is more and more convenient. The rapid develop-
ment of mobile Internet and the massive growth of mobile phone users, making the
various functions of APP are constantly changing in a rapidly changing network
environment. The development and replacement of the network security issues
have also received more and more people’s attention [1, 2]. Under the premise of
ensuring network security,, social platforms such as Twitter, Facebook, micro-blog
and WeChat, etc., have also rapidly emerged in the rapidly developing cyberspace
and gradually developed from a single web-based terminal to a dual-platform based
on mobile terminals.
∗ Guoqin Chang; Laboratory of Intelligent Computing and Application Technology for Big
Data, Henan University of Science and Technology, Luoyang, Henan, 471023, China, E-mail:
[email protected]
† Hua Huo – Corresponding author; School of Software, Henan University of Science and Tech-

nology, Luoyang, Henan, 471023, China, E-mail: [email protected], [email protected]

CTU
c FTS 2018 325
 Neural Network World 4/2018, 325–344

Many of the comment information and real-time short texts contain the per-
sonal sentiment and tendencies of users, which are of great significance for users
to personalize recommendations, interest circle division, network public opinion
monitoring, and privacy protection [3]. How to use the computer technology to
acquire and analyze these comments in the emotional information has attracted
many experts and scholars to research. The information has attracted the com-
peting research of many experts and scholars, which involves many fields such as
artificial intelligence, natural language processing, data analysis and mining [4].
The main purpose of sentiment analysis is to process, extract, summarize and
analyze the information in the text through different methods, so as to infer the
emotion and viewpoint expressed by the author of the text, and divide the emo-
tional tendency of the text through the subjective information contained in it. Text
sentiment analysis can be divided into 3 tasks, including sentiment information ex-
traction, sentiment information classification and sentiment information retrieval
and induction [5]. Sentiment information classification is one of the important
tasks of text sentiment analysis, and is also the focus of this paper.
Among them, Sina micro-blog has become the mainstream social platform with
more than 300 million users. It is through the mutual attention between users,
sharing real-time information, commenting on micro-blog content and other ways
to socialize. But since micro-blog released the text words provisions shall not be
more than 140 characters, the text length is short, the structure is different [6],
the sentiment content in the text is relatively less. By analyzing the characteristics
of this short text, this paper proposed a method of short text sentiment analysis
in micro-blog bases on machine learning, which the proportion of emotional words
and improves the accuracy of text sentiment analysis through the improvement of
feature selection and weight calculation algorithm.
The specific contents of this paper are organized as follows: The second part
mainly introduces the related work; the third part introduces the methods used in
this paper and explains the proposed algorithm for feature extraction and weight
calculation in detail; the fourth part gives the experimental results and analysis.
Finally, the full text is summarized.

2. Related work
Natural language processing has extensive research scope, which includes sentiment
analysis, machine translation, text classification, semantic analysis, etc. [7–9]. As
the practicality of sentiment analysis has gradually become a research hotspot in
recent years. Text sentiment analysis, also known as opinion mining, has become
one of the hot topics in the field of natural language processing and data mining due
to its practicality. Its main research methods can be divided into two categories,
one method is based on the sentiment knowledge, another method is based on
feature classification [10].
The method based on the sentiment knowledge is to use the existing sentiment
knowledge, such as the sentiment lexicons, domain dictionary, etc. to classify the
text. Turney et al. [11] put forward a method based on point mutual information
for the co-occurrence of words in the corpus and use the method to make senti-
ment judgments on the words. Sentiment words are the basis of text sentiment

326
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

analysis [12], building sentiment lexicons was also very important research content.
Baccianella et al. [13] used General Inquirer (GI) to construct an sentiment lexi-
cons, Gyamfi et al. [14] used the MPQA corpus [15] to establish an sentiment seed
lexicons and implemented an expansion of the sentiment lexicons in conjunction
with WordNet. The construction of sentiment lexicon gradually transforms from
relying completely on artificial addition to automatically adding new sentiment
words through various methods [16–20].
The method based on feature classification is to select a large number of features
that can represent the sentiment of the text and to classify the texts by means of
statistics or machine learning methods such as Naive Bayes, K-proximity, support
vector machine, etc. Nowadays, most of the paper are also used in this method [21–
24], and these methods have also been successfully applied to different fields [25].
Pang et al. [26] classify the movie reviews, they used the similarity of sentences as
a feature and training classification was conducted through NB and SVM that the
experiments show classification based on reviews is more difficult than classification
based on facts text. In the subjective and objective classification of English and
French, Toprak et al. [27] used words, part of speech and lexical information of
features, then used SVM as a classifier that the experiments proved that the recall
of lexical information features to improve subjective and objective classification
Better effect. Moraes et al. [28] used TF-IDF and GI as the feature extraction
algorithm respectively and then classified used Pang et al.’s data sets with classifiers
as NB, SVM and ANN, experiments proved the ANNs were found to be superior to
unbalanced data sets good expressiveness. At the same time, sentiment analysis has
also begun to gradually develop from coarse to fine-grained [29–31]. Fink et al. [32]
conducted sentiment analysis using method of machine learning which subjective
and objective classification using coarse-grained (sentence poles) and sentiment
classification using fine-grained (clause or part-of-speech). Shi et al. [33] made
use of the correlation of words and part-of-speech information to make sentiment
analysis of hotel reviews and achieved good results. However, the method based
on feature classification doesn’t take into account the influence of sentiment words
and there is still much room for improve the methods in feature extraction.
In addition, due to the continuous development of deep learning in various
fields [34], the research of sentiment analysis gradually begins to tend to unsuper-
vised classification [35–38]. Such methods are quite difficult and of great research
significance also can save labor, but the method is not mature enough yet and the
classified accuracy is relatively low which can’t be used to the application now.
The development of the network has promoted the large-scale growth of data.
The small language of foreign languages and the national language in Chinese have
begun to occupy a place in the network. For the protection and development of
various languages, the textual sentiment analysis of this small language has also
attracted many eyes of scholars [39–42].
Due to the existing methods, there are still some defects [43]. Therefore, we
combining the method based on sentiment knowledge and the method based on
feature classification, and proposed a text sentiment analysis method based on ma-
chine learning, which applies the sentiment lexicon to feature extraction to achieve
better classification results.

327
 Neural Network World 4/2018, 325–344

3. Fine-grained sentiment analysis method for

short text in micro-blog
According to the disadvantages in existing methods for sentiment word has less
proportion in feature and low weight, a new method of sentiment analysis for short
text in Micro-blog is proposed, increases the proportion of sentiment words in
feature extraction and weight calculation, the main process was shown in Fig. 1.

Fig. 1 Sentiment analysis process.

3.1 Text preprocessing

Since all the sentiment information was contained in the text, in order to ex-
tract features and analyze, first of all, the text should be preprocessed the text

328
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

to remove the excess interference information and retain useful information. Text
preprocessing is the basic and important link in sentiment analysis, and the quality
of preprocessing is very important for sentiment classification and information pro-
cessing in the later stage. The contents of text preprocessing include data cleaning,
word segmentation and removing stop word.

3.1.1 Data cleaning

As one of the most representative social networks nowadays, micro-blog has a
huge amount of users and a large variety of comments on it. This is also due to
the fact that this feature leads to the unique style of the texts in the Micro-blog
which are often not in conformity with the changing forms Grammatical norms.
However, these distinctive short texts contain a large amount of author’s sentiment
information. Due to the social nature of Micro-blog, these texts also include replies,
references and sharing. Although these contents can express the user’s emotions
to some extent, many redundant and redundant information and noise are also
involved in the extraction and influence to the theme and content. For example:

The content between the two wells of the # Champion not only dynasty endless
# and # Hengda seven consecutive # in the sentence belongs to Micro-blog’s topic
and the content expresses the joy of victory. The emotions expressed by users in
comment 1 were consistent with the topics, while the contents of comment 2 were
opposite to those expressed in the topics. Therefore, these topics do not fully
represent the emotional tendency of the user, which play a role of interference and
should be deleted.
For interference items similar to the hot topic existing in the data set, they
should be cleaned up in order to lay a good foundation for the extraction of the
later emotional information, and delete the contents as shown in Tab. I.

3.1.2 Word segmentation and removing stop word

After data cleaning, the text in the data set be preserved are strong availability, but
in order to achieve better classification effect and need to have text processing in
further, that is word segmentation and removing stop words. The objective was to
completely remove the redundant information in the data set as much as possible
and to retain the most valuable content.
Word segmentation: Both Chinese and English are very important parts of nat-
ural language. There are some similarities between the two languages but there
are still many differences. English words are separated by a space, but there is no
such structure in Chinese. We need word-level fine-grained sentiment analysis and

329
 Neural Network World 4/2018, 325–344

Delete Content Before Cleaning After Cleaning

Delete Topic 文章内容。这是荣耀手机#极限少 这是荣耀手机，她们去鸟
女养成营#，她们去鸟巢滑行 巢滑行

(This is Rongyao mobile phone # (This is Rongyao mobile

limit girls develop camp #,They go phone, They go to the
to the bird’s nest to slide) bird’s nest to slide)
Delete Mentioned
Person @三木摄影全球旅拍用镜头诠释西藏 用镜头诠释西藏的神秘；
的神秘；
(Use the lens to interpret
(@ Miki global photography trip the mystery of Tibet;)
Use the lens to interpret the mys-
tery of Tibet;)
Delete Reply 倍感欣慰// 蒙牛是很差劲 倍感欣慰

(Feel happy //Mengniu is very bad) (Feel happy)

Delete Source 湖 北 大 学 的 研 究 生 廖 可 富 ，1年 申 湖北大学的研究生廖可
请16项发明专利（武汉晚报）。 富，1年申请16项发明专
利。
(Liao Kefu, a graduate student at
Hubei University, applied for 16 in- (Liao Kefu, a graduate
vention patents in one year(Wuhan student at Hubei Univer-
Evening News).) sity, applied for 16 inven-
tion patents in one year.)
Delete Link 今天，给大家支两招 今天，给大家支两招

https://mp.weixin.qq.com (Today, give you two

(Today, give you two strokes strokes)
https://mp.weixin.qq.com)

Tab. I Data cleaning content.

one of the very important step is to carry out Chinese word segmentation. Chinese
word segmentation involves grammar, statistics and other fields of knowledge and
there are some difficulties, but now have a lot of more mature word segmentation
system. Through the screening and comparison, this paper uses NLPIR Chinese
word segmentation system (also known as ICTCLAS2016), this system has higher
accuracy in Chinese word segmentation and has better recognition effect on collo-
quial vocabulary and network new words recognition. The segmentation effect is
as follows:

330
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

Removing stop words: In Chinese there are some functional words that make
the text fluent, such as function words, prepositions, quantifiers and so on. These
words usually neither have practical meaning nor emotional information, but exist
in a large amount in the text, at the same time, there are also some extra infor-
mation such as Roman symbols, mathematical characters and punctuation, which
we collectively refer to as stop words. There are many different types of stop word
list, we integrate the existing stop words by screening and contrast, and retain
some adverbs which have some influence on the emotion of the text. The complete
segmentation data set to remove stop words, data preprocessing part is completed.

3.2 Feature extraction

In the process of sentiment analysis, feature extraction is also a very important
link. If all the words appear directly as the characteristics of sentiment analysis,
the data size is huge and the classification time is long. Therefore, we need to select
words with strong emotional tendency as the features and calculate the appropriate
feature dimensions, which can effectively avoid the over-fitting caused by excessive
feature dimensions and improve the efficiency and speed of the classification.
There are many existing feature selection methods, such as point mutual infor-
mation, information gain, chi-square test and so on. Comparing the advantages
and disadvantages of various methods, point mutual information without consider-
ing the influence of redundancy features and in the information gain all categories
can only use the same features, but the chi-square test is to avoid these shortcom-
ings, so the final choice is improved and the improved algorithm is used to select
features of the chi square test.
Chi-square test, also known as CHI statistics, reflects the degree of association
between feature words and categories. The formula is as follows:

N (AD − BC)2
CHI(t, c) = . (1)
(A + C)(B + D)(A + B)(C + D)
Among them t represents the characteristic term, c represents the category, N
represents the total number of texts, A, B, C, and D represent as shown in Tab. II.

Belongs to Category c Does not Belong to

Category c
Inclusion Feature t A B
Does not Include Feature t C D

Tab. II Formula representation.

Since N, (A + C), (B + D) are all constant in Eq. 1, then Eq. 1 can be simplified
as follows:

(AD − BC)2
CHI(t, c) = .
(A + B)(C + D)
However, since the chi-square test does not calculate the frequency of the feature
words in the statistical process, it is easy to select the words with the lower word

331
 Neural Network World 4/2018, 325–344

frequency and in the original method the influence of the sentiment words is not
taken into consideration. For these problems, we join the word frequency and
sentiment words into the algorithm that improved chi-square test as follows:

N-CHI(t, c) = CHI(t, c)f (t).

The improved algorithm was named N-CHI and its scores was used as the basis
for selecting the feature words. CHI(t, c) is the basis scores of the feature words,
and f (t) is the weight function of the feature words. The main contents of the f (t)
are as follows:


 2 t = SentimentWord & f > 1
1.5 t = SentimentWord & f = 1

f (t) =

 1 t 6= SentimentWord & f > 5
0.5 t 6= SentimentWord & f ≤ 5


where f is the frequency of the feature word. The algorithm adds some weight
to the original CHI algorithm, increases the score of the sentiment word as much
as possible, makes the sentiment word more easily become the feature word and
reduces the probability of the CHI algorithm selecting the low word frequency so
as to improve the classification accuracy rate of the later. The improvement effect
as shown Fig. 2.

Fig. 2 N-CHI improvement effect.

It can be seen from the figure that the number of sentiment words contained
in the improved algorithm is significantly higher than that of the no improved

332
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

algorithm in different dimensions. We choose the words with higher scores as the
features and the choice of their feature dimensions is also very important which
number is too much or too little will affect the accuracy rate. The specific content
of the feature data selection will be discussed in the fourth section.

3.3 Weight calculation

The weight refers to the value of the feature. This value shows the correlation
between this feature and all kinds of types, the commonly used weight calculation
methods are frequency, TF-IDF and so on. Because the TF-IDF algorithm is
easy to implement, not only the word frequency but also the influence of the anti-
document frequency are taken into account. Therefore, the TF-IDF algorithm is
improved and the improved algorithm is taken as the weight. For a document D
consisting of K categories, the formula for TF-IDF is as follows:
 
Dk
Wik = tfik · idfik = tfik · log +x .
dik + x
Among them Dk represents the total number of documents in the k-th category,
tfik represents the number of times that the feature word ti appears in document
Dk , it means the word frequency, dik represents the number of documents in which
Dk contains the feature word ti , and a constant x is usually set in idfik to smooth
the function curve and prevent the case where the frequency is 0 in one category.
However, since the original algorithm still does not take into account the influence
of emotion words, modifiers and so on, this paper improves the TF-IDF algorithm
and names it as W-TF-IDF. Its purpose is to make it more suitable for the task of
emotional classification, which is distinguished from common text classification by
the following formula:
 
Dk
Wik = (tfik · (1 + f1 (ti )) + f2 (ti )) · log + 0.5 .
dik + 0.5
f1 (ti ) represents the feature words ti whether a sentiment words, if ti is senti-
ment words the value is 1, not the value is 0; f2 (ti ) represents the feature words
ti whether a modifier, if ti is the modifier the value is 1, not a modifier value is
0, and the smooth curve constant x is set to 0.5. In this way, the weight of the
sentiment words can be sufficiently improved and the influence of the modifiers on
the sentiment is also taken into consideration. The improvement effect is shown in
the following Tab. III.

TF- W-TF- Sentiment Modifier

IDF(pos) IDF(pos) Words
喜欢 (Like) 246.92 507.84 YES NO
今天 (Today) 142.39 142.39 NO NO
极其 (Extremely) 6.21 12.99 NO YES

Tab. III W-TF-IDF improvement effect

333
 Neural Network World 4/2018, 325–344

It can be seen from the table that the improved algorithm for sentiment words
and modifiers has a significant increase in its weight. In the meantime, in order to
make the calculation more convenient and rapid in the later training classification,
we also normalized the expression of the weights.
Regardless of the improvement of the feature selection algorithm or the weight
calculation algorithm, the complexity of the algorithm is increased to some extent
compared with the original algorithm, but there is not much difference in the
running time. The main purpose is to improve the text sentiment classification.
The final correct rate.

3.4 Feature vector and constructing classifier

Feature vector: The selected features and calculated weights are not directly used
as input into the classifier, so we store them in the feature vector with an n-
dimensional vector of text d as follows:

d = [(t1 , w1 , (t2 , w2 , (t3 , w3 , . . . , (tn , wn )].

ti represents the n-th feature of the text, wi represents the weight of the n-th
feature, which is often used to represent features in the text classification because
it is simple and easy to implement and does not consider the correlation and order
of the features.
Construct classifiers: There are many ways to construct classifiers, such as
method based on the statistical, methods based on machine learning, method
based on deep learning, which method based on the machine learning are also
included in the MaxEnt (Maximum Entropy), KNN(K-Nearest Neighbor), NB and
SVM(support vector machine) etc. NB is a classic probabilistic model algorithm
that determines the classification result by calculating the probability that text
d belongs to class c, p(d | c) is the conditional probability that text d belongs
to class c and p(c) is the prior probability of that class. This method has simple
logic, classification algorithm, mature and stable characteristics, combined with the
characteristics of text classification using NB constructing classifier and achieved
better classification results, and compared with other more mature algorithms, the
formula is as follows:

p(d | c) · p(c)
p(d | c) = .
p(d)

4. Experimental design and results analysis

In this part, we mainly introduce the process and result of the experiment in detail
and give a brief analysis of the experimental results. The data set and training
were processed in python on a desktop computer with a 3.3 GHz Intel Core i3 CPU
and 2 GB RAM.

334
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

4.1 Experimental design

4.1.1 Experimental data
We used the COAE 2014 (Sixth corpus of Chinese tendentious analysis and uation)
task 3 and task 4 corpus as the original data, the data set contains Micro-blog com-
ments that have been labeled positive and negative two sentiment Which contains
1138 positive reviews and 1193 negative reviews. After the heavy reprocessing, have
1048 positive reviews and 993 negative reviews including 6222 words are included.
We set 90 % of them as training set and 10 % as verification set.
For the sentiment lexicon, we have generated a new sentiment lexicon through
selections, additions and subtractions between the sentiment words in Hownet and
NTUD’s sentiment lexicon in Taiwan University. The new sentiment lexicon include
7164 positive lexicons and 12,060 passive lexicons.

4.1.2 Evaluation method

Often used in machine learning, natural language processing, information retrieval
and other fields of evaluation indicators are the following categories: accuracy,
precision, recall and F -measure. The accuracy rate is defined as the ratio of the
number of correctly classified samplers to the total number of samples in a given
test set classifier. Although it can determine the effect of classifiers to a certain
degree, it is not the most objective and effective. Therefore, In the experiment, we
use the precision (P ), the recall rate (R) and the F -measure (F ) as the evaluation
criteria. The calculation method is as shown in Eqs. 2, 3 and 4, and the meanings
are shown in Tab. IV.
TP
P = (2)
TP + FP
TP
R= (3)
TP + FN
2·P ·R
F = . (4)
P +R

Positive Negative
Be Judged as Positive t TP FP
Be Judged as Negative t FN TN

Tab. IV Evaluation index.

4.1.3 Selection the method of weight normalization

There are many weight normalization methods. The purpose of normalization is
to apply the weights to the classifier better. The correct choice of normaliza-
tion method has great influence on the efficiency and accuracy of data processing.
Species normalization method for experimental comparison, the method is as fol-
lows:

335
 Neural Network World 4/2018, 325–344

Method 1: Linear function normalization is the linear transformation of the orig-

inal data to the range of [0, 1], the method to achieve the original data is scaled,
where X ∗ is the normalized data, X is the original data, Xmax and Xmin , respec-
tively the maximum and minimum values of the original data set, the normalized
formula is as follows:
X − Xmin
X∗ = .
Xmax − Xmin

Method 2: The weight of a feature in a class is divided by the sum of the

weights of the feature in each class to normalize the feature weight to [0,1], taking
into account the effect of low word frequency. When the feature weight is greater
than the feature item the average value of all categories will be normalized weight
retention, otherwise 0, the normalized formula is as follows:
 Xi
 P k
Xi ≥ X
Xi∗ = Xi .
 k=1
0 Xi < X

Method 3: Since the final purpose is to judge the textual tendencies, the weight
is directly reduced to [−1, 0, 1] according to the propensity. X1 is the weight of
the feature t in the positive class, X2 is the weight of the feature t in the negative
class, The method is as follows:

X1
 −1
 X1 = 0 ∪ X 2
< 0.6
∗ X1
X = 0 0.6 ≤ X2 ≤ 1.5 .
X1
X2 = 0 ∪ X2 > 1.5

 1

Since we do not know the optimal dimension of the feature temporarily, we

choose 6000 as the feature dimension and then use the three normalization methods
to process the feature weight respectively. The obtained results are shown in the
following Tab. V.

P R F
Method One 0.91 0.90 0.90
Method Two 0.89 0.88 0.88
Method Three 0.89 0.89 0.89

Tab. V Comparison of normalization methods.

As shown in the above table, the first method has higher P, R and F. The
method retains the advantage of the sentiment word in the weight, while the second
method ignores the function of the sentiment word. When dealing with the low
frequency words, the range is too large, resulting in more weight is 0, although
the third method is more intuitive to make the weight close to the classification,
but the method ignores the weight of feature words size, and range selection has
some limitations, scalability is not strong. So this paper choice the method one as
a method for weight normalization.

336
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

4.2 Experimental results

In this paper we improved the algorithm of feature selection and weight calculation.
In order to verify its improvement effect we designed four sets of experiments: The
first group is choice the feature dimension; The second group is the improved
verification of N-CHI algorithm; The third group is the improved verification of
the W-TF-IDF algorithm; The fourth group is the effect of the two algorithms on
the experimental results. The main results and analysis of specific experiments
described in detail below.

Experiment 1: choice the feature dimension In the process of training clas-

sification the selection of the feature dimension always has been a very important
part. When the number of feature words is small, these feature can’t fully express
the sentiment expressed in the text, resulting in a low correctness rate. If the fea-
ture is too much, in the training process easily lead to over fitting. Through many
experiments, it is found that when the feature dimension between 5800-5900 the
precision is the highest and the best classification result can be achieved. There-
fore, we choose 5950 as the dimension of the feature, and the experimental result
is shown in Fig. 3.

Fig. 3 Choice the feature dimension.

Experiment 2: improved verification of N-CHI algorithm In order to

verify the effect of improved W-TF-IDF algorithm, the three contrast test design
in this experiment, three group experiments were performed using the TF-IDF as
the weight calculation: The first group using mutual information as the feature
selection method, named MI; The second groups using CHI as a feature selection

337
 Neural Network World 4/2018, 325–344

method, named CHI; The third groups using improved N-CHI as feature selection
method, named N-CHI. With the increase of dimensions, three methods of P, R, F -
measure experimental results are shown as follows in Fig. 4, 5, 6.

Fig. 4 Improved verification of N-CHI algorithm -P.

Fig. 5 Improved verification of N-CHI algorithm -R.

It can be seen from the above figure that this method is effective for the im-
provement of the method of feature extraction. Although the effect at lower lati-
tudes is slightly lower than that of CHI, with the increase of dimension, the N-CHI
algorithm is in P, R and F above point mutual information and CHI.

Experiment 3: improved verification of W-TF-IDF algorithm In order to

verify the improvement of the W-TF-IDF algorithm, three sets of comparative ex-
periments were designed in this experiment. Three sets of experiments all adopted

338
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

Fig. 6 Improved verification of N-CHI algorithm -F.

CHI as the method of feature selection: The first group uses the word frequency as
the weight calculation method, named TF; The second group uses TF-IDF as the
weight calculation method, named TF-IDF; The third group uses a modified W-
TF-IDF as the weight calculation method, named W-TF-IDF. The experimental
results are as follows in Fig. 7.

Fig. 7 Improved verification of W-TF-IDF algorithm.

It can be seen from the figure that the text is effective for the method of weight
calculation. Although the effect of promotion is not obvious, the values of P, R
and F are all improved.

339
 Neural Network World 4/2018, 325–344

Experiment 4: improved verification of comprehensive the algorithm

In this paper, the method of feature selection and weight calculation is improved.
In order to verify its improvement effect, a total of four sets of experiments were
designed: The first group without any improvement, named CHI-IDF; The second
group is a method that only improves the feature extraction; The third group is
a method that only improves the weight calculation; The fourth group made two
improvements to the text, named C-CHI-IDF, The experimental results are shown
in the following Tab. VI.

POS-P NEG-P F R F
CHI-IDF 0.86 0.92 0.89 0.88 0.88
N-CHI 0.89 0.94 0.91 0.90 0.90
W-TF-IDF 0.86 0.94 0.90 0.90 0.90
C-CHI-IDF 0.91 0.95 0.93 0.93 0.93

Tab. VI Comparison of experimental results.

It can be seen from the above table that the text achieves good results for the
improvement of feature extraction and weight calculation, both in terms of positive,
negative and overall precision.
The above experimental results are the average of multiple experiments. By
comparison, it can be found that the improvement of feature extraction and weight-
ing calculation is effective, the classification accuracy rate is significantly higher,
and the weighting improvement is better than that of feature selection The effect is
more obvious, and the improved algorithm has significantly improved the accuracy
of the positive class.

Experiment 5: Selection of classification algorithms Using different clas-

sification algorithms to construct the classifier will result in different classification
results. According to the classification task of text sentiment analysis, this paper
uses the improved C-CHI-IDF method to select naive Bayes (NB) which is more
suitable for text classification and KNN, and artificial neural network (ANN) to
construct the classifier, and compare the classification results, the classification
results are shown in Tab. VII.

POS-P NEG-P F R F
NB 0.92 0.94 0.93 0.93 0.93
KNN 0.90 0.94 0.92 0.92 0.92
ANN 0.91 0.95 0.93 0.93 0.93

Tab. VII Selection of classification methods.

According to the above table, there is no significant difference in the classifi-

cation effect between NB and KNN when the sample size is small. However, the
amount of data continues to increase. The classification effect of NB is significantly
higher than that of KNN, and it is better for classification effect of NB in positive.

340
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

The classification accuracy rate of ANN is not significantly different from the other
two, but it has better performance for the sample imbalance of ANN.

4.3 Experimental analysis

Through the above experimental results and brief analysis, we can see that the fine-
grained text sentiment classification method based on machine learning proposed
by the text is effective. In order to reflect the experimental results more intuitively,
we convert some data into the form of graphs. As shown in Fig. 8.

Fig. 8 Comparison of experimental results.

As can be seen from Fig. 8, we also found that the correct rate of the method for
the negative class is significantly higher than the correct rate of the active class,
although the number of negative class sentences is much lower than the active
class. This shows that words in negative sentences are more likely to be selected as
features, and in Chinese natural language, people are more likely to carry negative
words when expressing negative emotions.
In this paper, we used a general evaluation data set. Other sentiment analysis
methods of the data set and some evaluation results are compared with the results
of this paper. The comparison results are shown in Tab. VIII.

5. Conclusion
According to existing method of the feature extraction and weight calculation with-
out considering the influence for the sentiment words, this paper proposed a method
of fine-grained Chinese sentiment analysis for short text in Micro-blog which pro-
posed the N-CHI and W-TF-IDF two kinds of new algorithm based on improved
the algorithm of CHI and TF-IDF. This method effectively improves the accuracy
of text sentiment information classification and has certain universality.

341
 Neural Network World 4/2018, 325–344

POS-P NEG-P P
Literature [42] – – 0.900
Literature [6] Method 1 0.895 0.837 –
Literature [6] Method 2 0.919 0.886 –
COAE2014-sjtu 0.954 0.885 0.919
Bjut-coae2014 0.914 0.915 0.914
scool 0.769 0.800 0.785
WB-SA 0.892 0.914 0.903
hut 0.901 0.848 0.875
LEO-WH Run1 0.964 0.791 0.877
Medians 0.891 0.850 0.877
C-CHI-IDF 0.910 0.950 0.930

Tab. VIII Comparison of normalization methods.

However, this method has some limitations that does not consider the influ-
ence of the relationship between part-of-speech and words on affective information,
which will be the focus of the fourth part of this article.

Acknowledgement
This research is supported by National Natural Science Foundation of China under
the Grant 61672210 and supported by the Henan Research Program of Foundation
and Advanced Technology under the Grant 162300410183.

References
[1] ZKIK K., ORHANOU G., HAJJI S.E. Secure Mobile Multi Cloud Architecture for Authenti-
cation and Data Storage. IGI Global. 2017, 7(2), pp. 62–76, doi: 10.4018/IJCAC.2017040105.
[2] JAIN A.K., GUPTA B.B. PHISH-SAFE: URL Features based Phishing Detection System
using Machine Learning. In: Cjc, Csi, 2015.
[3] VElURU S., GUPTA B.B., RAHULAMATHAVAN Y., et al.Privacy Preserving Text Analyt-
ics: Research Challenges and Strategies in Name Analysis. Handbook of Research on Securing
Cloud-Based Databases with Biometric Applications. 2015, pp. 67–92, doi: 10.13140/2.1.
3017.3760.
[4] PIRYANI R., MADHAVI D., SINGH V.K. Analytical mapping of opinion mining and sen-
timent analysis research during 2000–2015. Information Processing and Management. 2016,
53(1), doi: 10.1016/j.ipm.2016.07.001.
[5] ZHAO Y.Y., QIN B., LIU T. Sentiment analysis. Journal of Software, 2010, 21(8), pp.
1834–1848, doi: 10.3724/SP.J.1001.2010.03832.
[6] YUAN D., ZHOU Y., LI R., et al. Sentiment analysis of microblog combining dictionary and
rules. In: Ieee/acm International Conference on Advances in Social Networks Analysis and
Mining, IEEE, 2014, pp. 785–789, doi: 10.1109/ASONAM.2014.6921675.
[7] HARRAG F., HAMDI-CHERIF A., El-QAWASMEH E. Performance of MLP and RBF
neural networks on Arabic text categorization using SVD Neural Network World. 2010,
20(4), pp. 441–459.
[8] BARIGOU F. Improving K-nearest neighbor efficiency for text categorization Neural Net-
work World. 2016, 26(1), pp. 45–66, doi: 10.14311/NNW.2016.26.003.

342
 Chang G., Huo H.: A method of fine-grained short text sentiment analysis. . .

[9] MAUTNER P., MOUCEK R. Processing and categorization of Czech written documents
using neural networks. Neural Network World. 2012, 22(1), pp. 53–66, doi: 10.14311/NNW.
2012.22.004.
[10] KUPKA J., TOMANOVA I. Some extensions of mining of linguistic associations. Neural
Network World. 2010, 20(1), pp. 27–44.
[11] PETER TURNEY M.L. Measuring Praise and Criticism: Inference of Semantic Orientation
from Association. ACM Transactions on Information Systems, 2003, pp. 315–346, doi: 10.
1145/944012.944013.
[12] KU L.W., LIANG Y.T., ChEN H.H. Opinion Extraction, Summarization and Tracking in
News and Blog Corpora. In AAAI-CAAW,, 2006.
[13] BACCIANELLA S., ESULI A., SEBASTIANI F. Multi-facet Rating of Product Reviews.
European Conference on Ir Research on Advances in Information Retrieval. Springer-Verlag,
2009, pp. 461–472, doi: 10.1007/978-3-642-00958-7_41.
[14] GYAMFI Y., WIEBE J., MIHALCEA R., et al. Integrating knowledge for subjectivity sense
labeling. Human Language Technologies: the 2009 Conference of the North American Chap-
ter of the Association for Computational Linguistics. Association for Computational Lin-
guistics, 2009, pp. 10–18, doi: 10.3115/1620754.1620757.
[15] WIEBE J., BRECK E., BUCKLEY C., et al. NRRC summer workshop on multi-perspective
question answering, 2002, doi: 10.1017/CBO9781107415324.004.
[16] WANG K.E., XIA RUI. An approach to Chinese sentiment lexicon construction based on
conjunction relation. Proceedings of the 14th China National Conference on Computational
Linguistics. Guangzhou, China: CCL, 2015.
[17] KRESTEL R., SIRESDORFER S. Generating contextualized sentiment lexica based on la-
tent topics and user ratings. Proceedings of the 24th ACM Conference on Hypertext and
Social Media. New York, NY: ACM. 2013, pp. 129–138, doi: 10.1145/2481492.2481506.
[18] LIANG JUN, CHAI YU-MEI, YUAN HUi-BIN, ZAN HONG-YING, LIU MIN. Deep learning
for Chinese micro-blog sentiment analysis. Journal of Chinese Information Processing. 2014,
28(5), pp. 155-61.
[19] HUANG M.L., YE B.R., WANG Y.C., CHEN H.Q., CHENG J.J., ZHU X.Y. New word
detection for sentiment analysis. Proceedings of the 52nd Annual Meeting of the Association
for Computational Linguistics. Baltimore, Maryland, USA: Association for Computational
Linguistics, 2014, pp. 531–541, doi: 10.3115/v1/P14-1050.
[20] ZHAO YANYAN, QIN BING, SHI QIUHUI, LIU TING. Large-scale Sentiment Lexicon
Collection and Its Application in Sentiment Classification. Journal of Chinese Information
Processing, 2017, 31(2), pp. 187–193.
[21] YANG Y., XU C., RRN G. Sentiment analysis of text using SVM. Lecture Notes in Electrical
Engineering (2012)138 LNEE, pp. 1133–1139, doi: 10.1007/978-1-4471-2467-2_134.
[22] SILVA N.F.F.D., HRUSCHKA E.R., JR E.R.H. Tweet sentiment analysis with classifier
ensembles. Decision Support Systems.. 2014, 66, pp. 170–179, doi: 10.1016/j.dss.2014.07.
003.
[23] AGARWAL A., XIE B., VOVSHA I., et al. Sentiment analysis of Twitter data. The Work-
shop on Languages in Social Media. Association for Computational Linguistics, 2011, pp.
30–38.
[24] HE FEIVAN, HE YANXIANG, LIU NAN, et al. A Micro-blogging Short Text Ori-
ented Multi-class Feature Extraction Method of Fine-grained Sentiment Analysis. Acta
Scientiarum Naturalium Universitatis Pekinensis, 50(1), (Jan. 2014), doi: 10.13209/j.
0479-8023.2014.013.
[25] ZHOU X, WANG S, XU W, et al. Detection of Pathological Brain in MRI Scanning
Based on Wavelet-Entropy and Naive Bayes Classifier. International Conference on Bioin-
formatics and Biomedical Engineering. Springer, Cham, 2015, pp. 201–209, doi: 10.1007/
978-3-319-16483-0_20.
[26] PANG, BO, LEE, et al. A sentimental education: sentiment analysis using subjectivity
summarization based on minimum cuts. Proceedings of Acl, 2004, pp. 271–278, doi: 10.
3115/1218955.1218990.

343
 Neural Network World 4/2018, 325–344

[27] TOPRAK C., GUREVYCH I. A sentimental education: sentiment analysis using subjectivity
summarization based on minimum cuts. Deft’09 Text Mining Challenge, 2009.
[28] MORAES R., VALIATI J.F, NETO W.P.G. Document-level sentiment classification: An
empirical comparison between SVM and ANN. Expert Systems with Applications. 2013,
40(2), pp. 621–633, doi: 10.1016/j.eswa.2012.07.059.
[29] ZIRN C., NIEPERT M., STUCKENSCHMIDT H., et al. Fine-Grained Sentiment Analysis
with Structural Features, 2011.
[30] GUXMAN E., MAALEJ W. How Do Users Like This Feature? A Fine Grained Sentiment
Analysis of App Reviews. Requirements Engineering Conference.IEEE, 2014, pp. 153–162,
doi: 10.1109/RE.2014.6912257.
[31] TEH P.L., PAK I., RAYSON P., et al. Exploring fine-grained sentiment values in online
product reviews. Open Systems. IEEE, 2016, pp. 114–118, doi: 10.1109/ICOS.2015.7377288.
[32] FINK C.R., CHOU D.S., KOPECKY J.J., et al. Coarse-and Fine-Grained Sentiment Anal-
ysis of Social Media Text. Johns Hopkins Apl Technical Digest. 2011, 30(1), pp. 22–30.
[33] SHI H., ZHOU G., QIAN P., et al. An unsupervised fine-grained sentiment analysis model for
chinese online reviews. International Journal on Information. 2012, 15(10), pp. 4277–4294.
[34] WANG S.H, LV Y.D., SUI Y., et al. Alcoholism Detection by Data Augmentation and
Convolutional Neural Network with Stochastic Pooling. Journal of Medical Systems. 2018,
42(1), pp. 2, doi: 10.1007/s10916-017-0845-x.
[35] HUANG F., ZHANG S., ZHANG J., et al. Multimodal Learning for Topic Sentiment Analysis
in Microblogging. Neurocomputing2017, 253(C), pp. 144–153, doi: 10.1016/j.neucom.2016.
10.086.
[36] BAECCHI C., URICCHIO T., BERTINI M., et al. A multimodal feature learning approach
for sentiment analysis of social network multimedia. Multimedia Tools and Applications2016,
75(5), pp. 2507–2525, doi: 10.1007/s11042-015-2646-x.
[37] CHEN F., GAO Y., CAO D., et al. Multimodal hypergraph learning for microblog senti-
ment prediction. IEEE International Conference on Multimedia and Expo. IEEE, 2015, 1-6,
doi: 10.1109/ICME.2015.7177477.
[38] PORIA S., PENG H., HUSSAIN A., et al. Ensemble application of convolutional neural
networks and multiple kernel learning for multimodal sentiment analysis. Neurocomputing,
2017, doi: 10.1016/j.neucom.2016.09.117.
[39] KAPUKARANOV B., NAKOV P. Fine-grained sentiment analysis for movie reviews in
Bulgarian. Proceedings of Recent Advances in Natural Language Processing. 2015, 9(7), pp.
266–274, doi: 10.1126/science.aaa8685.
[40] LUO YAWEI, TIAN SHENGWEI, YU LONG, TURGUN·IBRAHIM,
ASKAR·HAMDULLA. Sentiment Analysis of Uyghur Text for Fine-grained Opinion
Mining 2016, 30(1), pp. 140–148.
[41] AL-SMADI M., QAWASMEH O., AL-AYYOUB M., et al. Deep Recurrent Neural Network
vs. Support Vector Machine for Aspect-Based Sentiment Analysis of Arabic Hotels’ Reviews.
Journal of Computational Science, 2017.
[42] ZHANG XIAOMEI, LI RU, WANG BIN, WU DI, GAO JUNJIE. Subjective and Objec-
tive Classification of Micro-blog Based on Feature Fusion. Journal of Chinese Information
Processing. 2014, 28(4), pp. 50–57.
[43] TANG D., WEI F., QIN B., et al. Sentiment Embeddings with Applications to Sentiment
Analysis. IEEE Transactions on Knowledge and Data Engineering. 2016, 28(2), pp. 496–
509, doi: 10.1109/TKDE.2015.2489653.

344