=Paper=
{{Paper
|id=Vol-1452/paper10
|storemode=property
|title=The Text Network Analysis: What Does Strategic Documentation Tell Us About Regional Integration?
|pdfUrl=https://ceur-ws.org/Vol-1452/paper10.pdf
|volume=Vol-1452
|dblpUrl=https://dblp.org/rec/conf/aist/MurashovS15
}}
==The Text Network Analysis: What Does Strategic Documentation Tell Us About Regional Integration?==
https://ceur-ws.org/Vol-1452/paper10.pdf
The Text Network Analysis: What Does Strategic
Documentation Tell Us About Regional Integration?
A.Murashov1, O.Shmelev2
1 Yaroslav-the-Wise Novgorod State University
(murashov.andrey@mail.ru)
2 Kalashnikov Izhevsk State Technical University
Abstract. Values and attitudes towards the regional integration process of the
Russian political elites are considered as an indication of what regional integra-
tion (RI) tends to be and how it evolves over time. This paper suggests how to
systematically grasp and integrate elite’s attitude into the analysis of RI by
means of text network analysis. The text network analysis allows one to visual-
ize the meanings and agendas present within political manifests which are sup-
posed to reflect values and attitudes towards RI of the local political elite.
Keywords. igraph, political elite attitude, R, regional integration, regional strat-
egy, system of indicators, text mining, text network analysis (TNA)
1 Introduction
This paper is a part of PhD thesis aimed at constructing a so-called System of Indica-
tors of Regional Integration in Russia. Values and attitudes towards the regional inte-
gration process of the Russian political elites are considered as an indication of what
regional integration tends to be and how it evolves over time. One of the shortcom-
ings of conventional approaches is insufficient and unsystematic consideration of
political elite's attitude towards regional integration and decision-making in this field.
The question how to systematically grasp and integrate elite’s attitude into analysis
remains open.
In comparative politics measuring the attitudes of the political elite is often under-
taken either by expert surveys or by the analysis of political manifests [De Lombaerde
P. et al., 2011]. Our interest lies in researching the political manifests – a regional
strategy reflects opinion of a local authority, like a party manifest directly reflects
opinion of a political party.
Research questions are those related to monitoring and analyzing regional integra-
tion process. How do regions cooperate? What forms of integration emerge within the
selected regions? What problems / challenges does integration impose? Which indus-
tries are mainly affected by integration process? How do regions choose their region-
partners? What is beyond the choice?
78
2 Background
From theoretical perspective this paper is supposed to contribute to the investigation
of values and attitudes towards the regional integration process that are represented in
political manifests. This topic is covered in particular by:
(1) Comparative Manifesto Project (CMP) maintained by Manifesto Research
Group. Their purpose is to discover party stances by quantifying their statements and
messages to their electorate, method used is quantitative content-analysis [CMP,
2014];
(2) Leontief Center’s Study of Russian Regions’ Strategies aimed at, among other
things, building a classification of regional strategies based on their content, method
used is expert review and content-analysis [Zhikharevich B. et al., 2013];
(3) Philippe De Lombaerde from United Nations University, Institute on Compara-
tive Regional Integration Studies (UNU-CRIS) and his team who employing multi-
disciplinary approach in developing quantitative and qualitative tools to monitor re-
gional integration process [De Lombaerde P. et al., 2011].
3 Method
From methodological perspective this paper applies an approach which combines two
methods - comparative text-mining and graph analysis – “text network analysis”. The
text network analysis allows one to visualize the meanings and agendas present within
political manifests. This approach outputs a graph of relations between key terms
where each node represents a term and edges express logical associations between
terms.
Putting it in a general scenario of social networks, the terms are taken as people
and the segments of text as groups on LinkedIn or Facebook, and the term-document
matrix can then be taken as the group membership of people. Several notions of co-
occurrence have been used in the literature to group words [Saeedeh M. et al., 2010]:
document-wise/sentence-wise /window-wise/syntax-wise co-occurrence. We build a
network of terms based on their co-occurrence in the same text segments (paragraphs)
extracted from the documents in the course of expert review. There is an edge be-
tween two terms if they appear in the same text segments (paragraphs). The weight of
an edge is its frequency [Batagelj V. et al. 2002, Polanco X., 2006]. Such a network
(or conceptual map [Chernyak Е. et al., 2014]) visualizes logical associations between
concepts presented in the political manifests.
1. Establish text corpus and transform it
Data to analyze is regional strategies of socio-economic development as a central
and most capacious source of information about political elite’s views on regional
integration process. We are interested in 6 Russian regions situated alongside the
Moscow – St Petersburg transport corridor: Moscow, Moscow region, Tver’ region,
Novgorod region, Leningrad region, St Petersburg. Their strategies are studied. There
may exist a wide range of other official documents on regional integration but unfor-
79
tunately we are not able to cover all of them, so we decided to limit our sample by
regional strategies only.
Using Atlas.ti (qualitative data analysis software) we establish text corpus and re-
trieve those text segments (paragraphs) from the regions’ strategies which refer to
regional integration process, refine it in a specific way then (lemmatization, filter
stopwords, punctuation and numbers removing, etc.).
2. Explore text corpus (igraph & tm packages)
Text network analysis is performed with R [Yanchang Zhao, 2012], specifically,
with packages {igraph} and {tm} (provides functions for text mining). We build a
document-term matrix, after that, it is transformed into a term-term adjacency matrix,
where the rows and columns represent terms, and every entry is the number of con-
currences of two terms, after that, frequent words and their associations (fast-
greedy.community) are found from the matrix.
3. Visualization
Finally, we visualize the result by means of {igraph} package in R environment:
(1) plot the graph to show the relationship between frequent terms (graph.adjacency,
layout = layout.fruchterman.reingold, delete.edges), (2) dendrogram (dendPlot).
4 Results
First we review general graph statistics. Snapshot of the network metrics is in the
table following (tab.1). Volume of the strategies varies from 396 vertices for Moscow
region to 800 vertices for Tver’ region. Function assortativity.degree uses vertex de-
gree (minus one) as vertex values. The coefficient throughout the corpora is negative
suggesting that the connected vertices tend to have different degrees. Centralization is
a general method for calculating graph-level centrality score based on node-level
centrality measure. Novgorod region’s strategy is that one having most centralized
structure (centralization degree of 68% of its theoretical maximum). We also arrive at
the conclusion that there is a substantial amount of centralization in the Moscow re-
gion’s strategy. In general, the power of individual terms varies rather substantially,
and this means that, overall, positional advantages are rather unequally distributed in
each strategy. The global version of clustering coefficient (function transitivity) indi-
cates that the degree to which nodes in a graph tend to cluster together is relatively
low. This makes sense since we removed from the graphs singular edges for the sake
of simplicity (here we refer to a parameter n which is discussed below). Fastgreedy
algorithm identifies from 6 to 10 communities in the graphs with moderate modulari-
ty. As we can see from the table 1 the graphs are quite similar in terms of their math-
ematical conception. Much more insightful and interesting results come from analysis
of the networks’ content.
80
Table 1. Graphs’ key metrics1
Strategy
Parameter {igraph} function
SP M LO NO TO MO
Number of vertices vcount 737 463 490 491 800 396
Number of edges ecount 5039 2311 2517 2913 7296 1897
Assortativity assortativity.degree -0.25 -0.34 -0.32 -0.24 -0.31 -0.29
Transitivity transitivity 0.19 0.22 0.16 0.21 0.22 0.16
Average path length average.path.length 2.54 2.64 2.48 2.43 2.52 2.58
Graph density graph.density 0.019 0.022 0.021 0.024 0.023 0.024
Centralization Degree centralization.degree 0.49 0.41 0.52 0.68 0.48 0.68
Centralization Closeness centralization.closeness 0.54 0.48 0.53 0.67 0.50 0.61
Centralization Betweenness centralization.betweenness 0.30 0.30 0.27 0.55 0.19 0.39
Eigenvector Centrality Scores centralization.evcent 0.92 0.91 0.92 0.92 0.91 0.92
Diameter diameter 13 10 13 13 10 14
Number of communities (best split) fastgreedy.community 6 6 10 8 8 8
Modularity (best split) fastgreedy.community 0.40 0.49 0.35 0.38 0.32 0.38
To demonstrate some examples for applying the strategies to study regional inte-
gration the graphs following are built (fig.1). They are based on the strategy of St
Petersburg. The graph (fig.1,a) is crowded with many vertices and edges, it represents
most of the ideas we can find in the strategy. To simplify the graph we remove insig-
nificant terms. With function delete.edges, we remove edges which have weight less
than a certain value. To do it in our experiment we introduce a parameter n referring
to a number of text segments (paragraphs) where a certain term appears. After remov-
ing edges, some vertices become isolated and are also removed. The produced graph
is on fig.1,b. The interpretation is that we exclude from the scope of analysis most
rare and random concepts.
Let us set n equal to 8. The resulting graph on fig.2,a is crowded with many verti-
ces and edges, we can interpret it at some extent but we need to get more precise pic-
ture. We identify vertices whose removal increases the number of connected compo-
nents in the graph. They are: city, petersburg, development, etc. To simplify the graph
and find relationship between terms beyond the selected keywords, we remove major
articulation points (or alternatively those terms whose removal, we expect, will lead
to a result we are looking for) so that the layout is rearrange and new concepts and
links between them are revealed. We see that some of the articulation points are not
necessarily meaningful but just the highly frequent words carrying less meaning than
those with a moderate or low frequency and are thus not very valuable to explore.
1 SP = St Petersburg, M = Moscow, LO = Leningrad region, NO = Novgorod region, TO =
Tver’ region, MO = Moscow region
81
Fig. 1. Example of graph evolution (a – initial graph; b – truncated graph)
Next, we try to detect communities in the graph. Graph community structure is cal-
culated with the fastgreedy algorithm [Kincharova A., 2013]. The nodes that cluster
together (communities) are shown with the same color on fig. 2, indicating contextual
proximity of the terms used. The communities tell us that the local authorities focus
quite heavily on patterns of spatial development , unique role of St Petersburg and its
attractiveness for migrants, close association between the City and Leningrad region,
etc.
Fig. 2. Graph improvement by managing articulation points (a – initial graph; b – refined
graph)
We can also have a further look at which terms colocations are most frequent in
each strategy (fig.3). Parameter n tells us how many times the plotted collocations
appear. Parameter n is a lower bound of the frequency, that is, collocation «Moscow –
82
St Petersburg» appears not less than 33 times in different text segments within Tver’
region’s strategy. Each strategy mentions those regions more frequently which are
supposed to be their main partners.
Fig. 3. Most frequent terms colocations
Plots for Moscow region and Leningrad region suggest that the development of the
two regions is strongly connected to two largest Russian cities that they surround. For
instance, the Leningrad region strategy suggests strong spatial planning on territories
adjoined to St Petersburg. The Moscow region strategy repeatedly highlights im-
portance of Moscow and its agglomeration.
5 Concluding remarks
Given the work-in-progress style of our study, it has a number of weaknesses; among
them is a lack of the comparative perspective. A solution might be to build a network
of text segments (paragraphs) based on the terms shared by them (two-mode net-
work). Alternatively, we would probably arrive at some interesting insights if could
put a network on another one to see joint term clusters and terms-outliers [Ermakov
A. et al., 2002]. We appreciate the more sophisticated models use phrases or n-grams
instead of words and this also could be a possibility to improve our analysis. These
basic tools implemented here are only the beginning of possibilities for applications
of TNA. It is worth noting that we were not meant to provide an extension of the base
technology of computer-supported TNA but suggest an example of its practical im-
plementation in social science.
The text-mining approach combined with graph theory appears to be a valuable
method for extracting elite’s attitude towards regional integration process from public
strategic documentation (allows us to access a large amount of textual material, re-
gional analysis may provide interesting input for text network analysis, etc.). Model-
ing the data using this method provided us with the specific insights on what local
authorities really focus on and how the strategies differ from each other.
83
References
1. Batagelj V., Mrvar A., Zaversnik M. Network analysis of texts. URL:
http://nl.ijs.si/isjt02/zbornik/sdjt02-24bbatagelj.pdf (25.03.2015).
2. Chernyak Е., Morenko Е., Mirkin B. Conceptual Maps: Construction Over a Text Collec-
tion and Analysis. Analysis of Images, Social Networks and Texts Communications in
Computer and Information Science Volume №436, 2014, pp.163-168.
3. De Lombaerde P. et al. The Regional Integration Manual: Quantitative and Qualitative
Methods. Routledge, London, 2011.
4. Ermakov A., Pleshko V. Informatization and information security of enforcement officials.
XI International scientific conference. Conference proceedings, Moscow, 2002, pp. 343-
347.
5. Kincharova A. Application of community detection algorithms for sociological investiga-
tion of blogs: results of a piloting study. URL:
www.hse.ru/data/2013/06/10/1283702757/dzh.pdf (25.03.2015).
6. Manifesto Project Database. URL: https://manifestoproject.wzb.eu/
(25.03.2015).
7. Polanco X., San Juan E. Text data network analysis using graph approach. Vicente P.
Guerrero-Bote. I International Conference on Multidisciplinary Information Sciences and
Technology, Oct 2006, Merida, Spain. Open Institute of Knowledge, vol. 2, pp.586-592.
URL: https://hal.archives-ouvertes.fr/hal-00165964 (25.03.2015).
8. Saeedeh M. et al. A Comparative Study of Word Co-occurrence for Term Clustering in
Language Model-based Sentence Retrieval. Human Language Technologies: The 2010
Annual Conference of the North American Chapter of the ACL, pp. 325–328, Los Ange-
les, California, June 2010. URL: http://www.aclweb.org/anthology/N10-1046
(25.03.2015).
9. Yanchang Zhao. R and Data Mining: Examples and Case Studies. Academic
Press, Elsevier, 2012.
10. Zhikharevich B., Zhunda N., Rusetskaya O. Proclaimed and actual priorities of regional
and local authorities: approaches to reveal and compare // The Region: Economics and So-
ciology, 2013, №2, pp. 108 – 132.
84