=Paper=
{{Paper
|id=Vol-3351/paper15
|storemode=property
|title=Sentence Level Fine-grained Emotion Computation Based on Dependency Syntax Improvement Dictionary
|pdfUrl=https://ceur-ws.org/Vol-3351/paper15.pdf
|volume=Vol-3351
|authors=Xiaoli Feng,Fangyuan Ju,Haotong Hou,Jiakai Peng,Yuwei She,Fulian Yin
|dblpUrl=https://dblp.org/rec/conf/aiotc/FengJHPSY22
}}
==Sentence Level Fine-grained Emotion Computation Based on Dependency Syntax Improvement Dictionary==
https://ceur-ws.org/Vol-3351/paper15.pdf
Sentence Level Fine-grained Emotion Computation Based on
Dependency Syntax Improvement Dictionary 1
Xiaoli Feng, Fangyuan Ju, Haotong Hou, Jiakai Peng, Yuwei She, Fulian Yin*
Communication University of China, Beijing, China
Abstract
With the rapid development of Internet technology, emotion analysis has been widely used in
various fields. However, the different types of emotion corresponding to multiple attribute
words in text and the complex syntactic structure bring great challenges to the existing text
emotion analysis methods. This paper proposes a sentence level fine-grained emotion compu-
ting approach based on dependency syntax improvement dictionary. Firstly, an attribute quad
extraction algorithm is constructed based on the dependency relationship. Then we analyze
the syntactic structure of the text, and study the influence of the combination mode of sen-
tence pattern, inter-sentence relation, degree adverb and negative adverb on the sentence
emotion. Lastly, a fine-grained emotion computing algorithm is designed to calculate the
emotion value of sentences. Experiments on collected micro-blog data set show that our
method compared with the original dictionaries, the F1 value of emotion classification reach-
es 81.37%.
Keywords
dependency syntax, emotion dictionary, fine-grained, emotion analysis
1 Introduction
Text emotion analysis, also known as opinion mining, refers to the analysis of the subjective text
with emotion color, mining the emotion tendencies contained in it, and dividing the of emotion atti-
tudes [1], which involves artificial intelligence, data mining and other fields [2]. Emotion analysis ini-
tially begins as an analysis of words with emotion [3], for instance, “beautiful” is a word with a posi-
tive meaning color, while “ugly” is a word with a negative meaning color. In 2010, Brendan
O’Connor et al. [4] applied emotion analysis to public opinion texts on Twitter, a foreign social net-
work platform, confirming that there was a strong correlation between voters’ emotion tendencies ex-
pressed on the Internet and poll results. Emotion analysis gradually attracted people’s attention and
was applied in many fields. For example, in the field of e-commerce, merchants can master users’ sat-
isfaction with related products and predict future sales through the analysis of product reviews [5]. In
the media field, the video platform can understand the audience's preferences and reasonably arrange
the broadcast time through the comment analysis of TV and films [6]. In addition, text emotion analy-
sis is also widely used in government public opinion monitoring [7], brand monitoring [8], stock mar-
ket prediction [9] and other aspects, which is of great significance and value.
At present, a series of studies on emotion analysis have been conducted abroad, but due to the dif-
ferences in grammar, sentence pattern and language habit between Chinese and English, there are still
many limitations when these studies are applied to the Chinese field [10]. Therefore, domestic schol-
ars are also actively exploring Chinese emotion analysis. Binary emotion analysis, that is, the user’s
emotion attitude is only composed of good and bad two extremes. Multivariate emotion analysis in-
troduces neutral emotion category. However, human emotion expression is complex. Emotions are not
limited to good, bad and neutral emotions, but also include more emotion states [11], such as anger,
panic, disgust, fear, happiness and other finer levels of division. Emotion classification based on emo-
AIoTC2022@International Conference on Artificial Intelligence, Internet of Things and Cloud Computing Technology
EMAIL: *yinfulian@cuc.edu.cn (Fulian Yin)
© 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
106
tion dictionary refers to the use of emotion dictionary to obtain the emotion value of emotion words in
the document, and then determine the overall emotion tendency of the document through weighted
computation [12-13]. However, due to the limited capacity of the emotion dictionary, scholars will
expand the emotion dictionary. At present, there have been researches on multi-category emotion
analysis in China, most of which use the method of expanding the emotion dictionary with large-scale
corpus. Hu et al. [14] proposed to use seed words in WordNet to generate an emotion dictionary con-
taining both positive and negative adverbs. Cruz et al. [15] proposed to use random forest algorithm to
expand the emotion dictionary. Ding et al. [16] summarized the steps of emotion computation in emo-
tion dictionary as marking emotion words, processing emotion reversal words, processing transition
words and summarizing emotion score. Although these researches can reduce the influence of the lag
of the emotion dictionary to a certain extent, their accuracy is low. At the same time, this method is
not suitable for emotion analysis of text with complex sentence structure and multiple attribute words.
Popescu et al. [17] identified candidate attribute words by calculating the Pointwise Mutual Infor-
mation (PMI) between candidate words and commonly used attribute words, but the selection of
benchmark words in this algorithm had a great influence on experimental results. Ku et al. [18] use
Term Frequency-Inverse Document Frequency (TF-IDF) to identify attribute words, but this model
ignores the extraction of infrequent words. Lu et al. [19] combined dependency syntax with domain
emotion dictionary for emotion computing, but the granularity of subjects selected in the experiment
was relatively coarse, so it was impossible to judge the performance of the algorithm in fine-grained
view content mining. Therefore, it is of great significance to extract attribute words and analyse sen-
tence structure in text emotion analysis.
In order to solve the problem of different emotion of multiple attribute words and complex sen-
tence structure, this paper proposes a sentence-level fine-grained emotion computing model based on
dependency syntax modified emotion dictionary to conduct text tendency analysis and emotion classi-
fication. This paper defines an extraction algorithm of emotion word quad in text based on dependen-
cy relation and expands the emotion dictionary. In addition, we define propensity shift rules and fine-
grained emotion computing algorithms to achieve fine-grained analysis of text emotion, in order to
obtain the polarity of text emotion, emotion classification and emotion value. Then the users’ views
and attitudes can be more accurately known, laying a foundation for subsequent applications in vari-
ous fields.
2 Our approach
This paper proposes a method to improve the emotion dictionary by using dependency syntax, and
carries out a more detailed emotion analysis of sentences combined with the study of sentence struc-
ture. The research in this paper is divided into three parts, namely, the expansion of the emotion dic-
tionary based on dependency syntax, fine-grained emotion computing, and experimental and result
analysis. First, attribute words, emotion words, negative adverbs and degree adverbs are extracted
from the corpus using dependency syntax, semantic dependency and part of speech. Then, the similar-
ity between the unknown emotion words in the quad and the seed words in the emotion dictionary is
calculated, and the unknown words are extended to the emotion dictionary. According to the complex
syntactic structure of Chinese text, the sentence tendency transfer rules are defined, and the effects of
sentence patterns, sentence relationships, degree adverbs and negative adverbs combination patterns
on sentence tendency are studied. Finally, this paper also designed the corresponding emotion data
computation algorithm, which is used to classify the text emotion and judge the emotion tendency.
The scope of this paper is shown in Figure 1, and we will introduce each module in detail.
2.1 Basic emotion dictionary construction
Emotion dictionary is the basis of emotion analysis. To carry out fine-grained emotion computa-
tion, it is necessary to select a dictionary containing both the emotion
polarity of words and the emotion classification of words. Therefore, this paper chooses emotion
vocabulary ontology database of Dalian university of technology as the basic dictionary. Ontology da-
tabase describes a word or phrase from multiple perspectives, including word category, emotion cate-
107
gory, emotion intensity and polarity. It mainly includes seven emotion broad categories, namely, joy,
like, anger, sadness, fear, disgust, surprise and twenty-one emotion small categories.
Figure 1 The architecture of this paper.
Negative adverbs are a kind of special adverbs. In most cases, when negative adverbs modify
emotion words, the actual emotion expressed will be contrary to the emotion of the word itself, and
then change the overall emotion color of the text. However, when negative adverbs modify degree
adverbs, the emotion intensity will be reduced. In this paper, fifteen commonly used negative adverbs
are selected through manual screening to obtain a negative adverb dictionary, and their emotion
weight is -1.
Degree adverbs can deepen or weaken the emotion expressed. When degree adverbs modify the
emotion words, the emotion intensity will be expanded by a certain multiple. This degree adverb
dictionary comes from the CNKI.com dictionary base, which divides degree adverbs into 6 grades,
namely, super, most, very, relatively, slightly and lack, and assigns certain weights to these 6 grades.
Examples of partial degree adverbs are shown in Table 1.
Table 1 Examples of DEGREE ADVERB DICTIONARY
Grades Degree adverbs Weight Number
super excessively, too 3 30
most extremely, unduly 2.5 69
very much too, how 2 42
relatively not big, more 1.5 37
Slightly a bit, a little 1 29
lack not that, poorly 0.5 12
2.2 Syntactic dependency-based emotion dictionary expansion
When emotion dictionary is used for emotion analysis, it is usually limited by the dictionary capac-
ity. The words that are not in the dictionary cannot participate in emotion computation, thus affecting
the accuracy of analysis. In addition, this method cannot take into account the emotion inconsistency
of multiple attribute words in the text.
Therefore, word extraction rules based on dependency relationship are designed to extract the
quads of attribute words, emotion words, degree adverbs and negative adverbs in the text to solve the
problem of multiple attribute words in the text. At the same time, the representative words are select-
ed from the emotion dictionary as the seed words, and the similarity computation between the un-
known emotion words and the emotion seed words in the quad is carried out to obtain the emotion at-
tributes of the unknown words, so as to realize the real-time extraction and emotion computation of
the unknown emotion words. Through threshold comparison and classification, unknown emotion
words are added to the existing emotion dictionary to obtain the expanded emotion dictionary, and the
subsequent text fine-grained emotion computing is carried out.
108
2.2.1 Syntactic dependency-based emotion dictionary expansion
Chinese text usually consists of multiple clauses to form complex sentences, and there may be de-
scription objects in each clause, which are called attribute words. For example, in the sentence “I real-
ly like my new mobile phone, take pictures very clearly, and play games do not stuck, but the battery
is a little small” have multiple attribute words, namely, “mobile phone”, “take pictures”, “play games”,
“battery”, and “like”, “clear”, “not stuck”, “small” are the modifiers corresponding to the four attrib-
ute words, namely emotion words.
In addition, the emotion of text is not only determined by emotion words, but also affected by neg-
ative adverbs and degree adverbs. Therefore, this paper proposes the construction of attribute as the
core of the quad, for each attribute word in the sentence to construct its “attribute word, emotion word,
negative adverb, degree adverb” quad.
Dependency syntactic analysis is a common technique in the field of natural language processing,
which can determine the dependency relationship between the syntactic structure of a sentence and
the words in the sentence. The result of dependency syntactic analysis can be expressed as a tree
structure, where stands for node and each word in the sentence is associated with a node. And stands
for directed edge, indicating that there is a dependency relation between two words.
In this paper, we define the extraction rules of quad for the common five types of dependency rela-
tions. Among the relations, the subject-verb (SBV) relation refers to the relationship between the actor
and the action. Verb-object (VOB) relation refers to a pairing relationship between a verb and an ob-
ject. ATT (attribute) relation refers to the relationship between a modifier and a centre word, modifier
is attribute. ADV (Adverbial) relation also refers to the relation between the modifier and the central
word, the difference is that the modifier is adverbial. COO (coordinate) relation refers to the relation-
ship between two or more concepts under the same concept [20].
2.2.2 Candidate word emotion judgment and emotion dictionary expansion
After word extraction of the text, the corresponding quad of each attribute can be obtained. The
words that contain or may contain emotion colors in the quad are called candidate emotional words.
When using the emotion dictionary to calculate emotion, it is usually necessary to use large-scale cor-
pus to expand the emotion dictionary in advance. If unknown words are encountered in the computa-
tion, it is impossible to know whether the words contain emotion, which will affect the emotion com-
puting. Therefore, this paper constructs an algorithm to determine the emotional attributes of un-
known words, which can directly calculate the emotional attributes of unknown words in the quad,
and obtain the emotion intensity, emotion category and emotion polarity
2.2.2.1 Emotion seed words selection
In the emotion judgment of unknown words, the method of calculating the association between
words is generally adopted to judge the similarity between unknown words and known words, so as to
determine their emotion characteristics. The candidate words with high similarity to the selected seed
words are more likely to be emotion words, or are more likely to be emotionless words. Seed words
are not only the basis for generating candidate emotion words, but also the benchmark words for judg-
ing algorithms.
2.2.2.2 Word embedding
Word embedding refers to mapping a single word into a vector. It is a common technique in the
field of natural language processing and has been widely used in text emotion analysis and similarity
computation. Word vector is the product of language model training. Commonly used word embed-
ding techniques include Word2vec, Glove, BERT and so on. BERT is an efficient tool to represent
words as vectors of real numbers and has made great achievements in the field of natural language
109
processing. BERT used MLM to pre-train the bi-directional Transformers to generate deep bi-
directional language representation. Therefore, Bert is used to obtain the word vector in this paper.
2.2.2.3 Cosine similarity
Before the text emotion analysis, the selected seed words are trained as word vectors. During the
sentence analysis, the unknown emotion words in the obtained quads are trained as word vectors to
calculate the similarity between seed words and unknown words. Cosine similarity is usually selected,
which measures the similarity between two vectors by calculating the cosine of the angle between
them, and can be used to calculate the cosine distance between two vectors. The higher the value is,
the more similar the two words are.
2.2.2.4 Candidate word emotion judgment and emotion dictionary expan-
sion
In the judgment of candidate words, words, while are more likely to be neutral words. Therefore,
after the emotion computation of candidate words and seed words, the candidate words are firstly
classified, and then their emotion intensity and polarity are calculated. Candidate word emotion judg-
ment and emotion dictionary expansion can be divided into the following five steps:
• Step 1: Train word vectors for candidate and seed words. Firstly, the candidate emotion words
in the text are obtained by using the quad matching rules, and then the seed words and candidate
words are trained as word embedding vector to form the seed word vector set SeedList .
• Step 2: Calculate similarity. The similarity of the selected seed word vector set
SeedList = {SeedWord1 , SeedWord 2 ,..., SeedWord 7 } and the emotion word vector to be added is
calculated.
• Step 3: Judge the emotion category of word. In the computation results of similarity of emotion
values between candidate words and various seed words, the three largest similarity values of
this kind of seed words are selected to calculate the average value, and the result is the
similarity between the word and this kind of emotion. If the similarity value is greater than 0.85,
the emotion category of the word is the emotion category corresponding to the maximum
similarity value, or it is classified as no emotion words and does not participate in the emotion
computation.
• Step 4: Capture the emotion intensity and polarity of words. After the candidate words are
classified, the seed words closest to the unknown words are extracted. The emotion intensity is
the intensity of the unknown words, and the polarity is the polarity of the unknown words.
• Step 5: Expand the emotion dictionary. The emotion category, emotion vocabulary and emotion
intensity of the above algorithms are successively written into the table to obtain the expanded
emotion dictionary, which is used as the emotion dictionary of subsequent emotion computation.
Finally, the expanded emotion dictionary used for fine-grained emotion analysis in this paper
mainly consists of four parts: emotion word, emotion category, emotion intensity and emotion polarity.
For a comment text, this paper can carry out syntactic analysis and extract the quads according to the
defined extraction rules. Then the similarity between the unknown words and the known emotion
words is calculated to obtain the emotion attributes of the unknown words, and the subsequent fine-
grained emotion calculation is carried out.
2.2.3 Fine-grained emotion computing model
The fine-grained emotion computing model proposed in this paper mainly includes tendency trans-
fer rules and emotion synthesis calculation. The architecture of fine-grained emotion computing mod-
el as shown in Figure 2.
110
Figure 2 The architecture of fine-grained emotion computing model.
2.2.3.1 Tendency transfer rules
In addition to obtaining the emotion of words, we also need to obtain the value of text tendency
transfer rules. First determines the sentence patterns of the text for exclamatory sentences, rhetorical
sentences to compute the weight of sentence pattern. Then, the sentences containing multiple clauses
are decomposed to obtain each clause, which is matched according to the relationship between clauses
in the tendency transfer rules to obtain the weight of the relationship between clauses. Finally, the
combination pattern matching of degree adverbs and negative adverbs is carried out to obtain the
weight of the combination relation of each clause. Due to space limitation, specific weight calculation
rules are not given in this paper.
2.2.3.2 Emotion synthesis calculation
In this paper, we use the extended emotion dictionary combined with the sentence tendency trans-
fer rules to conduct a fine-grained emotion analysis of chinese sentences, including the calculation of
emotion values and the judgment of emotion categories. The bottom-up comprehensive emotion cal-
culation is carried out for the Chinese text from word to sentence, and the emotion classification is
carried out after the emotion calculation value is obtained.
Based on the characteristics of corpus, this paper combines “joy” into “like” emotion, and the
emotion category is subdivided into six granularities: like, anger, sadness, fear, disgust and surpriser.
Each emotion granularity forms a unique emotion dictionary. When calculating the emotion value, the
existing emotion words in the text are compared with the constructed emotion dictionary. The com-
prehensive emotion value of sentence is calculated according to the emotion intensity of the corre-
sponding emotion words in the emotion dictionary, the weight of sentence patterns, the relationship
between sentences, the combination mode of degree adverbs and negative adverbs. The specific calcu-
lation steps are as follows:
• Step 1: Get the emotion intensity value. The emotion words were matched with the ontology
database, if the emotion words were in the ontology database, the emotion value seni is directly
obtained; if not, the similarity calculation of seed words was used to obtain the emotion
intensity value.
• Step 2: Get the emotion attribute value. For the degree adverbs and negative adverbs in the
word quad, the combination pattern of degree adverbs and negative adverbs is used to obtain the
weight. The calculation formula is as follows:
E (W ) = NA × seni (1)
Where E (W ) is the emotion attribute value, and NA is the combination weight of negative adverbs
and degree adverbs that modify the emotion word.
111
• Step 3: Calculate the emotion value of clause. The emotion value of each clause in the complex
sentence is calculated to judge whether there is an inter-sentence relation in the clause and get
the inter-sentence relation weight of the clause. The calculation formula is as follows:
E ( S j ) = i =1 E (Wi ) × T j
n
(2)
where E ( S j ) is the emotion value of the clause, T j is the inter-sentence relation weight, and E (Wi ) is
the emotion value of the quad in the clause.
• Step 4: Calculate the emotion value of complex sentence. When calculating the emotion value
of the whole text, the emotion value of each clause should be added and the influence of
sentence patterns should be considered. As shown in the following formula:
E ( S ) = i =1 E ( Si ) × Ri
n
(3)
where Ri is sentence pattern weight.
For any micro-blog text, various emotion values of the text can be obtained by the above emotion
calculation formula.
3 Experiments
In this paper, python is used to write crawler code, and 3272 texts are extracted from micro-blog.
After word segmentation and removal of stop words, the text content is transferred from traditional to
simplified and from uppercase letters to lowercase letters, and then manual emotion annotation is car-
ried out on the text, and emotion category and emotion tendency are marked. In order to reduce errors
caused by manual annotation, this paper recruited five college students with experience in annotation
to annotate corpus emotion. For a certain micro-blog text, if all annotated emotions are the same, take
them as text emotions, or the text is not included in the emotion calculation. In order to verify the ef-
fectiveness of the method proposed in this paper, the following three methods are designed for analy-
sis:
• Method 1: Emotion classification by using the basic emotion dictionary, namely, the Emotion
Ontology Database of Dalian University of Technology, degree adverb dictionary and negative
adverb dictionary.
• Method 2: Tendency transfer rules and basic emotion dictionary, negative adverb dictionary and
degree adverb dictionary are used to emotion classification.
• Method 3: Dependency syntactic improved dictionary and Tendency transfer rules are used to
emotion classification.
In this paper, joy and like emotion categories are combined as like, and commonly used F1 value
as evaluation metric of the model. Meanwhile, evaluation metrics with obvious improvement are
shown in bold in Table 2.
Comparation the results between method 1 and method 2, it can be concluded that the emotion
dictionary computing model with the tendency transfer rule has better performance in the overall
calculation accuracy and various emotion accuracy than the emotion vocabulary ontology database
model alone. The effect of sentence patterns, sentence relationships, degree adverbs and negative
adverbs on emotion was taken into account. By comparing the results of all kinds of emotion
classification, the accuracy of “like”, “sadness”, “fear”, and “surprise” is significantly higher than that
of “disgust” and “anger”.
For method 3, it can be concluded by observing the experimental results, the overall accuracy of
the emotion classification method based on the dependency syntactic modified emotion dictionary
constructed in this paper combined with the tendency transfer rules is much higher than the other two
methods. And for all kinds of emotions, the accuracy of emotion judgment has been improved to
varying degrees. Since the word embedding method does not consider the meaning of the words
themselves, words with similar context have similar word vector embedding, and there are errors in
the calculation of word similarity. Although this paper adopts the method of in-class average to judge
the emotion category and reduce the accidental error, it is not absolutely accurate to judge the emotion
of unknown words. In addition, the choice of seed words also affects the experimental results.
112
Therefore, the addition of the emotion calculation part of unknown words does not improve the
evaluation of all kinds of emotions in the model. The accuracy of “like” and “disgust” is higher than
that of “shock” and “anger”.
To sum up, the inclusion of sentence structure into the emotion analysis of microblog text
improves the emotion evaluation metrics of multiple categories to a certain improvement compared
with simply use the emotion vocabulary ontology database. After adding emotion words based on
dependency syntax extraction and calculating similarity to expand the emotion dictionary. Compared
with only join the tendency transfer rules in the evaluation metrics has the obvious improvement,
presents the better emotion classification effect, this model is verified in micro-blog this fine-grained
analysis of the feasibility of emotion.
Table 2 EMOTION CLASSIFICATION RESULTS OF THE THREE METHODS
F1
Category
Method 1 Method 2 Method 3
like 0.7177 0.7057 0.8590
anger 0.4696 0.4926 0.6188
sadness 0.7980 0.7644 0.9280
fear 0.8163 0.8208 0.9596
disgust 0.5573 0.6148 0.7524
surprise 0.5284 0.6899 0.7646
Average F1 0.6479 0.6814 0.8137
4 Conclusion
In view of the different types of emotion corresponding to multiple attribute words in the text and
the influence of complex syntactic structure on emotion analysis research problems. In this paper, we
propose a fine-grained emotion computation based on dependency syntax improvement dictionary
which can help mine user views and text values in texts and the validity of the model is verified.
In the next step, we can consider adding the part of network hot words to increase the coverage of
the emotion dictionary. The more complete the emotion dictionary, the better the effect of emotion
analysis. In addition, the text with unclear emotional expression also needs further research.
5 Acknowledge
The work was supported by the National Key Research and Development Program (No.
2021YFF0901705, 2021YFF0901700); the State Key Laboratory of Media Convergence and Com-
munication, Communication University of China; the Fundamental Research Funds for the Central
Universities; the High-quality and Cutting-edge Disciplines Construction Project for Universities in
Beijing (Internet Information, Communication University of China).
6 Reference
[1] T. Wang and W. Yang, “Review of Text Sentiment Analysis Methods,” Computer Engineering
and Applications. Beijing. China, vol. 57, pp. 11–24, April 2021.
[2] W. Hong and M. Li, “A review: Text sentiment analysis methods,” Computer Engineering &
Science. Changsha. China, vol. 41, pp. 750–757, April 2019.
[3] H. Vasileios and R. M. Kathleen, “Predicting the Semantic Orientation of Adjectives,” 35th An-
nual Meet. Association for computational linguistics and 8th Conf. Association for computational
linguistics European, pp. 174–181, 1997.
[4] O'Connor. Brendan, B. Ramnath, R. Routledge Bryan and A. Smith Noah, “From Tweets to Polls:
Linking Text Sentiment to Public Opinion Time Series,” 4th International Conf. Weblogs and
Social Media USA, pp. 122–129, 2010.
113
[5] J. Zhong, W. Liu, S. Wang and H. Yang, “Review of Methods and Applications of Text Senti-
ment Analysis,” Data Analysis and Knowledge Discovery, Beijing. China, vol. 5, pp. 1–13,
March 2021.
[6] M. Lei and M. Zhu, “Applications of sentiment analysis in movie recommendation system.
Computer Engineering and Applications,” Computer Engineering and Applications. Beijing.
China, vol. 52, pp. 59–63, October 2016.
[7] Z. Ren, P. Zhang, Y. Lan, Q. Zhang, Y. Xia and Y. Cui, “Emotional Tendency Prediction of
Emergencies Based on the Portraits of Weibo Users———Taking "8· 12" Accident in Tianjin as
an Example,” J. Intelligence. Xi’an. China, vol. 38, pp. 126–133, August 2019.
[8] X. Xu, Z. Yu and R. Meng, “Consumers Interaction，Brand Appeal and Purchase Behavior un-
der the Milieu of New Media: An Analysis based on Interaction ritual Chains,” Commercial Re-
search. Heilongjiang. China, vol. 2018, pp. 24–32, July 2018.
[9] J. Griffith, M. Najand and J. Shen, “Emotions in the Stock Market,” J. Behavioral Finance. vol.
21, pp. 42–56, January 2020.
[10] S. Zhou, W. Qu, Y. Shi, X. Shi and Y. Sun, “OVERVIEW ON SENTIMENT ANALYSIS OF
CHINESE MICROBLOGGING,” Computer Applications and Software, vol. 30, pp. 1–5, March
2013.
[11] K. Yang, M. Wu, and T. Chen, “Generalized text sentiment analysis review,” J. Computer Appli-
cations. vol. 39, pp. 6–14, December 2019.
[12] L. Chen, C. Lee and M. Chen, “Exploration of social media for sentiment analysis using deep
learning,” J. Soft Computing. vol. 24, pp. 8187–8197, June 2020.
[13] D. Pan, J. Yuan, L. Lin and D. Sheng, “Deep neural network-based classification model for Sen-
timent Analysis,” 6th International Conf. Behavioral, Economic and Socio-Cultural Computing
USA, pp. 1-4, 2019.
[14] M.Hu and B. Liu, “Mining and summarizing customer reviews,” 10th International Conf.
Knowledge discovery and data mining USA, pp. 168–177, August 2004.
[15] F. L. Cruz, J. A. Troyano, F. Enríquez, F. J. Ortega and C. G. Vallejo, “‘Long autonomy or long
delay?’The importance of domain in opinion mining,” J. Expert Systems with Applications, vol.
40, pp. 3174–3184, June 2013.
[16] X. Ding, B. Liu and P.S. Yu, “A holistic lexicon-based approach to opinion mining,” Web
Search and Data Mining, 2008, pp. 11–12.
[17] A. M. Popescu and O. Etzioni, “Extracting product features and opinions from reviews,” Natural
language processing and text mining, Springer, London, pp.9–28, October 2007.
[18] L. W. Ku, Y. T. Liang and H. H. Chen, “Opinion extraction, summarization and tracking in
news and blog corpora,” AAAI spring symposium: Computational approaches to analyzing web-
logs, pp. 100-107, 2006.
[19] H. Lu, Z. Niu, N. Zhang, X. Sun and W. Liu, “A Model for Sentiment Classification of Chinese
Microblog Based on Parsing and Theme Extension,” Transactions of Beijing Institute of Tech-
nology. Beijing China, vol. 30, pp. 824–830, August 2014.
[20] A. He, “The relation of incompatible juxtaposition and opposition between concepts,” J. Tianshui
Normal University. Gansu China, vol.1984, pp.43–45, April 1984.
114