The paper proposes a method for constructing and usage of the semantic models (SM) for the purpose of continuous monitoring of public opinion, opinion mining (OM) for finding actual subtopics in the Internet message flow. We define semantic model within this work, as the subject domain model, which has the form of a directed graph, vertices of which correspond to concepts of the domain, and edges define relations between them. Semantic models make it possible to use the results of linguistic statistical analysis of texts (Text Mining) and the use of Information Extraction methods for texts from the Internet for opinion mining. While existing public opinion analysis projects are more focused on one-time (static) public opinion analysis on objects and phenomena, this paper proposes a method for automated construction and use of SM based on continuous monitoring of public opinion on the Internet. OM procedure consists of three steps: construction and clustering of the SM; selection of documents and sentiment definition of subtopics; visualization of results. SM construction using compactified horizontal visibility graph algorithm, usage of cluster analysis methods for determining relevant subtopics, estimation of proportion and tonality for individual subtopic in overall topical information flow are shown. As examples, the models of subject areas corresponding to: “One Belt, One Road”, “Nord Stream”, “Genetically Modified Organisms” topics are considered. Obtained results confirm that proposed method can be used for opinion monitoring in in various subject areas.
We define semantic model, as the subject domain model, which has the form of a directed graph, vertices of which correspond to concepts of the domain, and edges define the relations between them. Concepts can be events, processes, i.e. such a semantic model can be interpreted as a semantic domain map.
Information that is created by Internet users reflects public opinion on various issues and can be collected, analyzed by content monitoring systems and taken into account when planning the activities of companies, organizations, etc. Semantic models make it possible to use the results of linguistic statistical analysis of texts (Text Mining) and the use of Information Extraction methods contained in texts from the Internet for opinion mining. While the existing public opinion analysis projects are more focused on one-time (static) public opinion research on objects and phenomena, this paper proposes a method for automated construction and use of SM based on continuous monitoring of public opinion on the Internet. 2
As part of this work, authors propose a public opinion analysis based on natural language processing methods. This analysis is aimed at determining the attitude of the subject of monitoring public opinion to the chosen topic. One of the main objectives of the analysis of public opinion is classification of emotional coloring of the text (positive, negative or neutral).
Current works devoted to the analysis, extraction of opinions, moods (Sentiment
Analysis - SA, Opinion Mining - OM) note that this is a computer study of opinions,
people's attitudes to an object, a concept that can represent individuals, events or
topics [
The work [
In [
In [
In this paper, we propose method for constructing and using SM for OM tasks on the Internet, which involves three stages [6]: - construction and clustering of SM; - selection of documents and determining the tonality of subtopics; - visualization of results.
At the first stage:
- selection of an array of documents for the construction of CM;
- finding concepts; definition of SM connections by constructing a compactified
graph of horizontal visibility [
At the second stage is made:
- selection of documents corresponding to subtopics (subthemes) from the general information flow using queries; - determine their share in the total flow of documents; - the tonality of the documents of the relevant subtopics is determined. At the third stage of the subtopics with the tonality: - visualized on the map; - states are recorded in the database (DB) of the monitoring system for subsequent receipt of the dynamics of the results change over time.
Below are considered the main operations performed as part of these three stages. 3.1
Selection of an array of documents to build a semantic model. On the basis of a given monitoring object and topic, a request is formulated for sampling an array of documents.
Finding concepts. Documents included in the array are pre-processed, service information is removed, as well as stop words that do not carry a semantic load. Stemming can be performed (coercion of words to the base). Then, on the basis of taking into account the frequency of words in an array of documents, or using other well-known metrics, for example TFIDF, the most important, having the greatest weight of concepts are selected from the words of the array of documents [8].
Definition of SM links by constructing a graph of horizontal visibility. The
algorithm of compactified horizontal visibility graph (CHVG) [
1. On the horizontal axis, a number of nodes are marked, each of which corresponds to words in the order of appearance in the text, and weight numerical evaluations are laid on the vertical axis (visually, a set of vertical lines).
2. Build a traditional graph of horizontal visibility. At the same time, there is a connection between the nodes, if they are in “line of sight”, i.e. if they can be connected by a horizontal line that does not cross any other vertical line.
3. The network obtained in the previous step is compactified. All nodes with this word are combined into one node. All connections of such nodes are also combined.
The peculiarity of using the CHVG algorithm in this work is that its first two steps are performed separately for each sentence of the analyzed text. After that, the resulting network is compactified. In the process of developing the proposed method, a study was carried out on the construction of a SM for documents collected on the following topics: One Belt, One Road (OBOR is the initiative of the PRC on the New Silk Road); Nord Stream; GMO et al. (For more detailed analysis of topics and documents, see the results section). A fragment of the graph of the semantic model built for 28 concepts of the OBOR topic using the described algorithm is shown in Fig. 1.
Clustering graph SM. Given the relevance of the aspect level of the analysis of
opinions, after building a semantic model, its network structure is analyzed using graph
clustering algorithms— community detection (clustering graph, community
detection). In [
Among the well-known community search algorithms in the graph, we can
distinguish the Louvain algorithm [
The Leading Eigenvector [
Walktrap clustering algorithm [
Fragment of the results of clustering of SM graphs constructed for documents collected on the OBOR topic using the considered algorithms are shown in Table 1.
The first line of Table 1 shows the names of three algorithms that were used to cluster the SM. Columns 1–2 show the results of applying clustering algorithms to the SM based on the concepts selected by frequency of use in the documents, and columns 3–4 show the results of clustering SM based on the concepts selected using the TFIDF indicator. The columns of the fragment of the cluster table show the interrelated sets of concepts of the analyzed topic, which were found using three clustering algorithms.
Formation of queries corresponding to clusters. As a result of the clustering of the semantic model graph, the sets of the most connected graph vertices were found, corresponding to the identified clusters, i.e. sets of similar concepts. On the basis of these concepts in the general stream of documents characterizing the analyzed subject area, subtopics, aspects are highlighted. Experts in this subject area give names to these subtopics and formulate requests for subsequent selection using the information retrieval system of documents corresponding to subtopics from the general flow of documents characterizing the general subject area.
For the OBOR topic under discussion, on the basis of the clusters found, the experts formulated four subtopics and the corresponding queries, given in Table 2.
Louvain
Walktrap economy military world countries country state global war trade projects economic cooperation security political infrastructure investment development international support national
chinese people market ness million including billion summit ers
president government xi busi
debt beijing
leadindia china south asia power region pakistan relations influence us strategic east europe russia american trump japan africa part sea
prime minister foreign policy china's initiative silk road
belt project
prime minister foreign road belt initiativ silk
Louvain (TFIDF) india military power pakistan country us strategic might russia american political iran russian defense syria turkey
sri president people ment xi trump investment cpec port chinese governprojects
debt project billion bri south region relations indian xinjiang economic cooperation africa development beijing summit pacific taiwan
prime minister foreign japanese abe mahathir malaysia china's obor research global market
Leading eigenvector
(TFIDF)
Subtopic name
President of the People's Republic of China Xi Jinping about investment projects as part of the Belt and Road Initiative (BRI)
The attitude of India, Pakistan, the United States and other countries to the initiative BRI
The attitude of countries to the development of the southern region in the framework of the initiative BRI Prime Minister of Japan Shinzō Abe about the initiative BRI
At this, periodically repeated training monitoring system (repetition period from several hours to a day), implemented using the SM construction and clustering algorithms, ends. 3.2
Selection of subtopics documents from the general information flow using queries. From the general flow of documents generated by the search query, which characterizes the subject area, the documents of subtopics are selected using search queries, formulated by experts on the basis of clusters identified in the semantic model. For selected documents, each of the subtopics is determined by their share in the total flow of documents.
For the flow of documents generated by the search query to the content monitoring
system InfoStream (one-road) & (one-belt) & china, which characterizes the OBOR
topic, the names of the subtopics and the corresponding requests are given in Table 2.
The shares of the subtopic documents selected using the formulated requests from the
total flow, are shown in Table 3.
Determining the tonality of the subtopic documents. For the documents of each of
the identified subtopics, the tonality is determined - positive, negative, neutral based
on the analysis of the words that make up the documents relating to the topics. To
determine the tonality, the algorithms proposed in [
At this stage, the visualization of the found subtopics with tonality on the map is performed. Monitoring results are visualized in real time on a geographic map with reference to specific objects. For each subtopic identified in the general flow of documents, a chart is displayed indicating the name of the subtopic and the proportion of documents in the total flow attributable to this subtopic, as well as the proportion of documents in positive, negative and neutral tonality within the subtopics (table 3, figure 2).
Fig. 2.Visualization of monitoring results on a geographic map. For each of the subtopics found, the map shows the proportion of its documents in the total input stream and the tonality of documents.
The geographic map shows: the subtopics identified in the flow of input documents; share of documents on each subtopic; the tonality of the documents on the subtopics, as well as the dynamics of changes in results over time. The conditions found during the monitoring are recorded in the monitoring database for later obtaining the dynamics of the results change over time. The change in share of documents and tonality of the formulated subtopics (subthemes) of the main OBOR topic by week is shown in Fig. 3, Fig. 4 andFig. 5.
1 3 5 7 9 11 13 15
17
WEEKS
The operations performed in the three stages of the considered method are implemented using the tools of the Gephi software package (http://gephi.org), as well as using software tools developed in the programming language for statistical calculations R. The results obtained using the proposed method summarized in the next section.
9 11 WEEKS 13 15 17 0 - GMO - genetically modified organisms and some other topics.
To monitor public opinion onOBOR topic, an array of 1000 English-language documents (from 30.11.2018 to 07.25.2018) collected using the (one-road) & (onebelt) & china query using the InfoStream system was analyzed.
At the first stage, after sampling an array of documents and its preliminary processing, finding concepts (based on the frequency of use of terms, as well as on the basis of the TFIDF indicator), the corresponding SMs were built (Fig. 1). Clusters of concepts obtained using the Louvain, Leading Eigenvector and Walktrap algorithms based on the constructed SM are shown above in Table 1.
Based on the comparison of the clusters obtained, experts in this subject area can classify document subtopics, as well as requests for a selection of documents relating to the subtopics, determining their share and tonality (Table 2).
At the second stage, the share of documents of the subtopics selected using the formulated requests from the general flow is given in Table 3. Also, the tonality of the documents of the formulated subtopics selected from the general flow of documents of the OBOR topic are given there.
At the third stage, the visualization of the results is performed. A general view of the interface for visualizing the found subtopics, their shares in the general flow and tonalities is shown in Fig. 2. The dynamics of changes in the share of documents and the tonality of the formulated subtopics by the composition of the OBOR topic documents by week are shown in Fig. 3-5. Fig. 3 shows the change in the proportion of documents in the formulated subtopics (Subtopic 1 - Subtopic 4, four lower graphs) in the composition of the documents of the OBOR topic (Topic, upper graph) by week. Fig. 4 shows graphs of changes in the number of documents with a positive, negative and neutral tonality in the subtopic of the “South Region Development” by week. Fig. 5 shows plots of changes in the number of documents with a positive, negative and neutral tonality in the “India Pakistan US” subtopic by week.
In addition to the OBOR topic, the use of the proposed method of public opinion monitoring on the topics of Nord Stream, GMO and others was considered. For example, for the Nord Stream topic, on the basis of 1000 English-language documents (collected from 02/11/2018 through 08/18/2018) a corresponding SM was built, clusters were found and subtopics were formulated: merkel putin meeting (about Chancellor of Germanyt Merkel and Putin meeting); gas transit ukraine (on gas transportation through Ukraine); european security energy market (on the security of the European energy market), poland united states (on the relationship of Poland and the United States to the Nord Stream project). 5
A method for constructing and using SM for public opinion monitoring has been proposed, which includes three stages: the construction and clustering of SM; selection of documents and the definition of the tonality of subtopics; visualization of results.
Construction of the SM using the compactified horizontal visibility graph algorithm, use of cluster analysis methods for determining relevant subtopics, estimating the proportion and tonality of individual subtopics within the general topical information flow are shown.
Obtained results confirm the possibility of using the proposed method of monitoring public opinion in various subject areas.