1. Introduction

With the development of Internet technology, social network has become an important platform for people to obtain and exchange information (Liu & Li 2010). Network public opinion has become the main source of public opinion (Li & Zhang 2014) and the main platform for people to express their opinions on hot issues, social politics, etc. (Tran & Tran 2017). A large number of network public opinion information contains many valuable parts. But effective information cannot be efficiently obtained from mass data using ordinary analysis methods because of the large amount and high complexity of network public opinion information, but the emergence of data mining brings a new direction for analysis of network public opinion information. Network public opinion information can be analyzed and utilized through text mining (Guo et al. 2014). Liu (2010) proposed a method of hot spot detection of network public opinion. First, vector space model was introduced to express the text format, and then K-means algorithm was used to cluster the corpus. Finally, support vector machine (SVM) classifier was used to classify the text. Chen et al. (2016) put forward a collaborative filtering based network public opinion trend prediction method, analyzed its principle, constructed the framework of network public opinion trend prediction, and verified the effectiveness of the method through experiments. Yang (2017) adopted the improved TF-IDF algorithm and introduced the part-of-speech weight coefficient and the position weight of feature words to improve the clustering effect of feature words in network public opinion and better reflect text characteristics. Wang & Tang (2016) put forward a Hidden Markov Model based semantic orientation analysis method to analyze and classify the emotional orientation of network public opinion and found that the algorithm could effectively improve the efficiency and accuracy of network public opinion analysis. Emotional tendencies refer to people’s inner preferences and dislikes for certain objective things. Emotional orientation analysis refers to excavating users’ psychological attitudes towards characters or events through the analysis and induction of the subjective text content generated by the network. Information of network public opinion usually include people’s attitudes towards events. Through clear emotional analysis, people’s attitudes towards public opinion can be obtained, which can help guide public opinion and is conducive to social security and stability (Hao et al., 2015). Emotional orientation analysis needs to be based on a large number of data, and the era of big data makes it possible to analyze emotional orientation. Through text mining technology, natural language text can be processed to obtain specific information. In the present study, the emotional tendency in network public opinion was studied, pattern matching was combined with machine learning method, and the emotional inclination in the evaluation of Meituan hotel on the Internet was analyzed using PMML method to verify the effectiveness of the method.

2. Network Public Opinion and Network Security

Public opinion refers to the set of people’s attitudes and emotions towards a matter, which is a reflection of people’s ideological status. Because of the particularity of the Internet, there are also some characteristics of Internet public opinion.

1. Universal and anonymous. The spread of information on the Internet is very fast and extensive. The speech made by netizens can be greatly diffused in a very short period of time. Because of the anonymity of the network, freedom of speech is more prominent. The public is unscrupulous when making comments, which also makes the network public opinion more complex.
2. Freed and controllable. The Internet provides a platform for people to express their opinions. Internet public opinion has a higher degree of freedom than the traditional paper media; there the true and false information is difficult to distinguish. But the network public opinion is not absolute freedom. Certain management and control is very necessary. The correct guidance can also play a certain role in the direction of network public opinion.
3. Interactive and immediate. The network platform provides great convenience for people’s information exchange, not only between people, but also between the public and government and between the public and media. Moreover the immediacy of the network also leads to the immediacy of the network public opinion and the timeliness of speech transmission.
4. Large influence. Because of the above characteristics, the influence scope and extent of network public opinion are very huge. A hot event can spread to a great extent in a very short period of time under the influence of network public opinion, but the event will become more complex because of different feelings and attitudes. The true or false information will affect the judgment of the public, resulting in confusion.

Complex network public opinion information has brought great challenges to network security. Due to the high freedom and openness of the Internet, the public opinion on the Internet is chaotic. A large number of pornographic, reactionary or violent information is common on the Internet, and the Internet not only makes these garbage and harmful information have a rapid spread of shortcuts, but also makes the vast number of netizens directly and passively face the garbage information.

Besides, the immediacy and high influence of Internet public opinion also poses a great threat to social stability. Faced with hot social and political issues, netizens can make unrestricted speeches on the Internet platform. If people are willing to take advantage of the Internet, they can lead public opinion to an aspect that is not conducive to social stability and stir up public sentiment. If timely and effective control is lack of, it is easy to make the situation develop in a bad direction and form network emergencies, which can bring great negative effects to the country and society.

3. Emotional Tendency Analysis on Network Public Opinion Based on Text Mining

At present, emotional tendency analysis is based on statistics, machine learning and semantics. In this paper, pattern matching was combined with machine learning, and the process of pattern matching and machine learning (PMML) is shown in Figure 1.

In order to reduce the computational load, emotional tendency is divided into positive and negative regardless of the neutral tendency, which are marked as 1 and – 1, respectively. At the training stage, the text is manually annotated, and the keywords are extracted to establish a keyword library. Then key words which has the largest classification determination degree are used to constitute a pattern library.

In the classification stage, key words are extracted from the words which needed to be classified, and emotional patterns, i.e., pattern characteristics, are generated according to the established pattern. They are matched with the pattern library, and then the emotional tendency is calculated to get the pattern sequence. Finally, the intensity of the emotional tendency of the text is obtained by machine learning such as SVM.

3.3. Calculation of emotional tendency

3.3.1. Calculation of key word tendency

W1 was taken as the subject and HowNet emotional dictionary which is a natural language processing and research software from HowNet was taken as the basis in the calculation of key word tendency. Seed set with evident commendatory and derogatory emotional tendencies is selected, and the emotional tendency of new words is obtained according to the similarity between new words and words in the seed set. For word c₁ and c₂, the similarity is defined as S(c₁, c₂), and the distance between words is defined as D(c₁, c₂), then

(1)

$S(c_1,c_2)=\frac{\partial }{D(c_1,c_2)},$

M1 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ S({c_1},\,{c_2})\, = \,\frac{\partial }{{D({c_1},\,{c_2})}}, \] \end{document}

where ∂ stands for a controllable parameter and D is a constant.

Through the calculation of word similarity, semantic tendency measures can be obtained. The positive emotional paradigm phrase is defined as cp, and the negative emotional paradigm phrase is defined as cn; the number is M and L respectively. The formula for tendency P of word c is:

(2)

$P(c)=\frac{1}{M}{\displaystyle \sum _{m=1}^{M}S(c,c{p}_m)-\frac{1}{L}{\displaystyle \sum _{l=1}^{L}S(c,c{n}_l)}}.$

M2 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ P(c)\, = \,\frac{1}{M}\,\sum\limits_{m = 1}^M {S(c,\,c{p_m}) - \frac{1}{L}\,\sum\limits_{l = 1}^L {S(c,\,c{n_l})} }\,. \] \end{document}

The word is considered as commendatory if the value of P was positive; otherwise it is derogatory.

3.3.2. Calculation of tendency under different patterns

The tendency of different patterns is very different; therefore it should be considered separately.

Adjectives in E1 and E4 indicate emotional tendency, so the tendency of patterns is the tendency of adjectives.

E2, E5, E8 and E10 include adverbs, which will strengthen emotional tendency. The degree of adverbs is divided into three levels, strong (very and especially), medium (relatively and slightly), weak (a little and some), and they are given different emotional weights R(c).

(3)

$R(c)=\{\begin{array}{c}2,\mathit{\text{strong}}\\ 1,\mathit{\text{moderate}}\\ 0.5,\mathit{\text{weak}}\end{array}$

M3 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ R(c) = \left\{ {\begin{array}{*{20}{c}} {2,\ strong}\\ {1,\ moderate}\\ {0.5,\ weak} \end{array}} \right. \] \end{document}

In E3, if the verb is an emotional word and the modifier is an adverb of degree, then it is the same as E5. If the verb is not an emotional word, then the adjective means emotional tendency, the same as E1.

The emotional expression of E9 is more complex, and the same expression may have different emotional tendencies in different contexts. For example, the emotional tendency of “consume fast” is different in “the power of the electroplate consumes fast” and “cupcakes in the store consume fast and usually are sold out before afternoon”. Therefore the tendency of E9 needs to be marked manually.

Because the existence of negative words in E6 and E7 will change the polarity of emotional tendency, therefore it is necessary to change the polarity after acquiring emotional tendency. In addition, when the two negatives words appear at the same time, it means double negative.

3.4. SVM based classification method

After the value of each pattern is obtained, the pattern sequence vector can be obtained as well. According to the pattern sequence vector, the tendency of text can be obtained by classification based on machine learning. N-Gram feature extraction is still needed in classification, but pattern matching can greatly reduce the number of features needed and the amount of computation.

For linearly separable problem, there is a training sample set {(x₁, y₁) (x₂, y₂), …, (x_N, y_N)}, in which x_i ∈ X, y_i ∈ Y = {1, –1}. Solve a linear decision function g(x) = w · x + b, where w stands for weight vector and b stands for constant. Interval is maximized on the premise of satisfying y_i(w · x_i) + b] ≥ 0, i = 1, 2, …, N, then

(4)

$\{\begin{array}{c}\text{min }\frac{\Vert w\Vert }{2}\\ \mathit{\text{st}}.(y_i(w\cdot x_i)+b)-1\ge 0,i=1,2,\cdots ,N\end{array}.$

M4 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ \left\{ {\begin{array}{*{20}{c}} {{\rm{\min}} {\rm{\,\,\,\,\,}}\frac{{\left\| w \right\|}}{2}}\\ {st.{\rm{\,\,\,\,\,}}({y_i}(w \cdot {x_i}) + b) - 1 \ge 0,i = 1,2, \cdots, N} \end{array}} . \right. \] \end{document}

Solve it using lagrangian multiplier, then

(5)

$w(\alpha )=\frac{1}{2}{\displaystyle \sum _{i,j=1}^n{\alpha }_i{\alpha }_j{y}_i}{y}_j(x_i\cdot x_j)-{\displaystyle \sum _{i=1}^n{\alpha }_i},$

M5 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ w(\alpha)\, = \,\frac{1}{2}\,\sum\limits_{i,\,j = 1}^n {{\alpha _i}{\alpha _j}{y_i}} {y_j}({x_i} \cdot {x_j}) - \sum\limits_{i = 1}^n {{\alpha _i}} \,, \] \end{document}

where α stands for Lagrangian multiplier.

The final decision function is

(6)

$f(x)=\mathit{\text{sgn}}\left({\displaystyle \sum _{i=1}^n{\alpha }_i{y}_i(x_i\cdot x)+b}\right).$

M6 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ f(x)\, = \,{\mathop{\rm sgn}} \left({\sum\limits_{i = 1}^n {{\alpha _i}{y_i}({x_i} \cdot x)\, + \,b} } \right). \] \end{document}

For nonlinear problem, linear classification after non-linear conversion can be realized after slack variable ζ is introduced. The objective function is:

(7)

$\{\begin{array}{c}\text{min }\frac{{\Vert w\Vert }^2}{2}+C{\displaystyle \sum _{i=1}^N{\zeta }_i}\\ \begin{array}{c}{y}_i\left[(w\cdot x_i)+b\right]\ge 1-\zeta i=1,2,\dots ,N\\ \zeta \ge 0\end{array}\end{array},$

M7 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ \left\{ {\begin{array}{*{20}{c}} {\min {\rm{\,\,\,\,\,}}\frac{{{{\left\| w \right\|}^2}}}{2}\, + \,C\,\sum\limits_{i = 1}^N {{\zeta _i}} }\\ {\begin{array}{*{20}{c}} {{y_i}\left[ {(w \cdot {x_i})\, + \,b} \right]\, \ge \,1\, - \,\zeta {\rm{\,\,\,\,\,}}i = 1,\,2,\, \ldots,\,N}\\ {\zeta \,\, \ge \,\,0} \end{array}} \end{array}} \right.\,, \] \end{document}

where C stands for penalty factor.

Through solution based on Lagrangian multiplier, the final classification function is:

(8)

$f(x)=\mathit{\text{sgn}}\left({\displaystyle \sum _{i=1}^n{\alpha }_i{y}_{i}K(x_i\cdot x)+b}\right).$

M8 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ f(x) = {\mathop{\rm sgn}} \left({\sum\limits_{i = 1}^n {{\alpha _i}{y_i}K({x_i} \cdot x) + b} } \right)\,. \] \end{document}

4. Instance Analysis

5. Discussion and Conclusion

With the development of science and technology, the Internet has gradually become the main cradle of public opinion (Song et al. 2014). With the development and popularization of the Internet, network public opinion is playing an increasingly important role in society. More and more network public opinions has become social public opinion problems through the Internet. It is found that network public opinion contributes to more than half of social events, and positive public opinion can raise social concern and promote the solution of events. However, it is difficult to distinguish the true from the false, and there is harmful information in public opinions. Because of the high openness and freedom of the network, harmful information is difficult to be completely eradicated, which brings about a lot of network security problems. The endless harmful information exerts a subtle influence on people’s mood and thinking through various means and encourages the dissemination of improper speech. With its growth, its influence extends from the network to the reality, causing social panic and turbulence. Therefore, the influence of network public opinion can not be ignored, and the massive public opinion data information needs to be processed and analyzed using effective methods. Text mining technology is of great help to the analysis of network public opinion information. Network public opinion analysis includes sensitive problem mining, hot spot tracking, trend control and emotional orientation analysis (Wang & Liu 2011). Through the analysis of emotional tendencies, we can grasp the people’s real feelings and attitudes toward people or events, understand the people’s position in current politics, and make decisions that conform to public opinion (Luo 2014). Analysis of emotional tendency of hot news can guide public opinion, stabilize social order, and prevent bad comments.

In order to effectively analyze public opinion information on the Internet, this study proposed PMML. Firstly, a pattern library was established by extracting keywords. Then, keyword orientation and pattern orientation were calculated. Finally, emotional orientation classification was realized by SVM. In order to verify the validity of the method, this study classified the emotional orientation on the evaluation on “Meituan Hotel” on the Internet. As to feature representation methods, UniGrams, BiGrams and TriGrams were compared, and it was found that BiGrams method in combination with PMML had the best classification performance, with a forward accuracy rate of 86.75%. Hence BiGrams was used for text classification. The results showed that 35.02% of the 7892 reviews was positive, 64.61% was negative, and 29 texts (0.37%) were not classified. The results showed that the method had good performance in classifying emotional tendencies of the hotel reviews and the negative tendency was greater than the positive tendency. The findings is an alarm for hotel managers. They are urged to enhance hotel management and provide better services. The 29 unclassified texts indicated that there were some shortcomings in the method, which needs to be further studied, and also showed that the text mining was difficult.

Emotional tendency analysis of network public opinion information is of great significance to maintain network security and promote social stability. In this paper, a PMML based method was proposed to classify the emotional tendency of network public opinion. The experiments showed that the accuracy of PMML in combination with BiGrams feature representation method was 86.75%, which indicated that the method was feasible, has broad application prospects, and strengthen network harmony and security.