Content uploaded by Raghava Rao Mukkamala
Author content
All content in this area was uploaded by Raghava Rao Mukkamala on Dec 04, 2017
Content may be subject to copyright.
Chapter 11
Methodological Demonstration of a Text
Analytics Approach to Country Logistics
System Assessments
Aseem Kinra, Raghava Rao Mukkamala and Ravi Vatrapu
Abstract The purpose of this study is to develop and demonstrate a
semi-automated text analytics approach for the identification and categorization of
information that can be used for country logistics assessments. In this paper, we
develop the methodology on a set of documents for 21 countries using machine
learning techniques while controlling both for 4 different time periods in the world
FDI trends, and the different geographic and economic country affiliations. We
report illustrative findings followed by a presentation of the separation of
concerns/division of labor between the domain expert and the text analyst. Impli-
cations are discussed and future work is outlined.
Keywords Logistics ⋅Transport system evaluation ⋅Big data analytics ⋅Data
mining ⋅Text mining ⋅Machine learning
A. Kinra (✉)
Department of Operations Management, Copenhagen Business School,
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
e-mail: aki.om@cbs.dk
R.R. Mukkamala ⋅R. Vatrapu
Computational Social Science Laboratory, Department of IT Management,
Copenhagen Business School, Howitzvej 60, 2000 Frederiksberg, Denmark
e-mail: rrm.itm@cbs.dk
R. Vatrapu
e-mail: rv.itm@cbs.dk
R. Vatrapu
Mobile Technology Laboratory, Faculty of Technology,
Westerdals Oslo ACT, Oslo, Norway
© Springer International Publishing Switzerland 2017
M. Freitag et al. (eds.), Dynamics in Logistics, Lecture Notes in Logistics,
DOI 10.1007/978-3-319-45117-6_11
119
11.1 Introduction
The assessment and appraisal of transportation systems is a central activity related
to both national and regional policy making for passenger and freight movement. In
this regard, there are a fixed set of widely accepted core methods for economic
evaluation and appraisal for different project investments, e.g., related to transport
infrastructure for the movement of people and goods. Most of these methods that
are currently used for this purpose involve the calculation of benefit-cost ratios.
There is an inherent assumption in that such an evaluation and assessment process
will aid economic goals, and even foster development in the case of developing
countries by lowering the overall transportation and mobility costs of businesses and
private users. However, the cross-border movement of goods involves international
transportation, global supply chains, multinational businesses and users, where the
costs and benefits of transportation systems are not only highly variable, but also tied
to the broader global logistics and supply chain systems that embed these trans-
portation systems. The Multi National Corporation (MNE), and its global supply
chain then becomes a primary user and motor of economic growth in an era of
globalization and global integration of businesses. Tracking the business consider-
ations of MNEs, and how their assessments of the national and regional trans-
portation, logistics and supply chain systems play a role in these considerations,
becomes important for policy makers in order to increase international trade and to
attract foreign direct investments in almost all sectors of the economy.
Toward this end, this paper explores and exploits text analytics methods and
techniques from big data analytics to develop a methodology for information
scanning, extraction, and retrieval of the logistics-related measures preferred by
decision-makers of MNCs, and other firms involved in global supply chains. This
information is relevant for the benchmarking and evaluation of transportation and
logistics systems. We use text mining and text analytic approaches, and more
specifically machine learning techniques in order to develop this methodology. We
then test run this methodology on a global supply chain text corpus in order to
illustrate and provide some feedback on the methodology. Our initial results
illustrate that the methodology is rather useful, and it can be successfully applied to
other logistics and transportation applications. However, we can also relate to the
general observation that model specification is the Achilles’Heel of big data
analytics as it is easy to under- or over-specify the model without deep involvement
of domain experts and deep knowledge of domain-specific problems.
11.2 The Assessment of Country Logistics Systems
The assessment of country logistics systems is an emerging area and literature is
sparse. Though most of the existing work on the assessment of country logistics
systems can be related to the useful distinctions in existing location research (see
120 A. Kinra et al.
Beugelsdjik et al. 2010; Beugelsdjik and Mudambi 2013). Kinra (2015) traces the
main literature streams in this area to the international economics, economic
geography, and international business distinctions in existing location research, and
these literature streams are now summarized.
The economics-oriented literature generally adopts the stance that spatial vari-
ation is generated by the structure of national resource endowments, and trading
patterns. The main contributions in this stream come from Memedovic et al. (2008)
and Bookbinder and Tan (2003), among others. For this group of literature spatial
discontinuity generally occurs at national borders and is defined by national
logistics systems. The geography-oriented literature tends to adopt the stance that
connectivity and colocation effects at major trade interface points generate spatial
variation. The contributions of Rodrigue and Notteboom (2010) and Rodrigue
(2012) are prominent in this stream. For these contributors spatial discontinuity
occurs at major nodes: transit points, hubs, and gateways, and is defined by regional
logistics systems, and hence the focus on the ranking of regional logistics systems.
Finally, the international business-oriented literature tends to adopt the stance that
the managerial and organizational utility from locations generates spatial variation.
Min (1994) is one of the earliest to demonstrate this stance. Similarly Carter et al.
(1997), Kinra and Kotzab (2008) make important advances in this regard. For this
group of contributors, spatial discontinuity generally occurs by country, and is
defined by the country investment climate.
All these contributions seek to provide an evaluation of the national logistics
systems. However some issues are common in the literature. First, the literature is
not unanimous about the logistics-oriented determinants, or factors, of location.
Second, it does not clarify the role and type of information, or measures, that are
adopted in the assessments. The literature is therefore mute about the specific
spatial impedance information and information categories that are relevant for
decision-makers in country assessment, and it is indefinite about the relevance and
the importance of the considered location determinants and the missing informa-
tion. This is a problem because information about the foreign country is widely
recognized as a major contributor in the complexity under conditions of global-
ization (see McCann and Mudambi 2004), and global supply chains.
We make a first attempt to the identification and categorization of the relevant,
available information on country logistics systems in Kinra (2015). Apart from the
identification of 187 different types of relevant and essential information measures
that fall under 17 decision factors, the study establishes the main spatial transaction
cost categories under which the information is made available, but also concludes
on conditions of information-driven complexity for the ranking of country logistics
systems.
However these findings are only based only complex manual procedures for
discerning this information, and include observations from a limited dataset of
countries for a limited time period (2006–2008). Nor any implications are presented
for the ranking (assessment) of the different countries based on the spatial impe-
dance measures that have been identified. This study will go beyond these
11 Methodological Demonstration of a Text Analytics Approach …121
limitations by developing more automated techniques that are capable of handling
bigger information sets. This will take better into account factors such as longitu-
dinality of data for determining the general conditions of information-driven
complexity hypothesis, but also from the point of view of comparing different time
periods. The next section now provides a detailed description of this methodology
development, including the text mining and text analytic approaches that are
employed to develop the methodology for automating the information scanning,
extraction, and retrieval process.
11.3 Methodology
In this section, we describe the methodology that was developed to conduct big data
analysis of global supply chain document corpus. The corpus primarily contains 21
text documents describing transportation and logistics systems pertaining to 20
countries spanning over period from 2006–2014 years. In order to analyze the text
corpus, we have used both text mining and machine learning techniques as shown
in Fig. 11.1.
As a first step (Text Extraction) toward applying text mining/machine learning
algorithms on the text corpus, we have extracted texts from the global perspective
documents which are in Portable Document Format (PDF) using an open source
software component using Java Programming language. The size of a global
Fig. 11.1 Text analytics methodology of global supply chain document corpus
122 A. Kinra et al.
perspective document varies from 20 to 60 pages and text from each page has been
extracted separately for further processing. Unfortunately, it was not possible for us
to extract the PDF metadata such as links and bookmarks from the documents. The
extracted texts are tokenized by breaking them sentences and words by using space
as delimiting character between the words using Natural Language Toolkit (NLTK)
(Bird 2006) and Python programming language. The final output of this step ended
up in raw text containing approximately 464,000 tokens, which is used as an input
for further processing using text mining and machine learning algorithms.
11.3.1 Text Mining
As part of text mining we have applied three different techniques: Keyword
Analysis,Word Frequency Analysis, and Collocation Analysis as further described
below. In order to prepare the raw text for applying these techniques, we had to
clean and preprocess the raw text using Natural Language toolkit. As part of
preprocessing step, we further processed the raw text tokens by removing nonal-
phabetical characters, numerical and stop words that are high-frequency words like
the,to, also, and so on, which do not contribute much to the semantic meaning of
the documents. Even though the raw text contains 464 000 tokens, after the pre-
processing step, we ended up with approximately 300,000 tokens.
11.4 Keyword Analysis
As part of keyword analysis, we investigated which decision factors are more
predominant in the country descriptions. In order to investigate that, each factor
(e.g., Waterways) is supplemented with suitable search words (port, water, sea,
shipping, etc.) that are representative of that factor. Even though 17 decision factors
with suitable search words are adopted from Kinra (2015) initially, soon the list of
factors got expanded to 21 factors to include more interesting factors. One of the
challenges with the decision factors from Kinra (2015) is that the number of search
terms varied from factor to factor which made the cross comparisons of decision
factors among the countries difficult. Therefore, we have readjusted search terms by
dropping few search terms for some factors and by adding new search terms for
some other factors, thereby making the number of search terms equal to four for
every decision factor. Soon after finalizing the search terms for decision factors, we
used NLTK with Python programming language to conduct analysis to find out the
number of occurrence of each search term (or keyword) in each perspective doc-
ument for all the decision factors. The occurrence counts of search terms are
summed up to find out the keyword occurrence count for a decision factor. Based
on the keyword occurrence counts of decision factors, we were able to compare
which decision factors are predominant in the entire corpus and as well as in each
11 Methodological Demonstration of a Text Analytics Approach …123
country descriptions. In addition to that, we could also be able to compare decision
factors across country dimension such as which are the most prominent countries
for any given decision factor.
11.5 Word Frequency Analysis
Word frequency analysis is a method of automatically identifying the frequent
occurring words from a given text corpus, by using the term document matrix. In
order to compute the word frequencies, raw tokens from the preprocessing step are
further analyzed using NLTK to prepare term document matrix containing the
frequency of words across perspective document collection. We have computed the
term document matrix for the whole corpus as a single document and for each
individual perspective document as well. The construction of term document
matrices enables us to find most frequent words (e.g., 100 most frequent words)
with the word frequencies, which is used to generate word cloud for a given
document. Word clouds or tag clouds is a simple way of visualizing or highlighting
the most frequently used words from a given text document. As part of the analysis,
we have generated word clouds for each perspective document and also a word
cloud for whole text corpus by combining all the perspective documents into one.
Word frequency analysis and word clouds enabled us to get an overview of major
concepts/topics discussed in the documents.
11.5.1 Collocation Analysis
Collocations are expressions of multiple words, which commonly co-occur in the
documents and therefore a collocation is a sequence of words that occur together
unusually often in the text documents. Finding collocation expressions involves
standard statistics-based and linguistically rooted association measures against mere
frequency of word occurrence counts. The collocation analysis provides insights
about the documents by providing bigrams, trigram, and n-grams that contain
words, which co-occur in the documents. The collocation analysis for each per-
spective document as well as for whole data corpus was conducted using Natural
Language toolkit (Bird 2006) to find out bigrams and trigrams. It provided us with
intuition about the topics and concepts that are discussed in the country descriptions
of transportation and logistics. It also helped us notice the emerging trends (other
than decision factors) in the perspective documents, which further helped us process
the documents for emerging trends or factors.
124 A. Kinra et al.
11.5.2 Machine Learning and Text Classification
As part of the methodology, our plan is to apply machine learning algorithms on the
text corpus to perform text classification tasks. Text classification approach comes
under supervised machine learning and it can be defined as a process where
assigning a predefined category of labels to new documents based on probabilistic
measure of likely hood using a training set of labeled documents (Yang and Liu
1999). Out of several approaches available in text classification domain, we have
chosen a simple text classification method (Zhang and Li 2007) based on Bayes rule
that relies on a simple representation of documents using bag of words approach. In
this project, we have used Naïve Bayes classification method to classify the
extracted text pieces from the perspective documents.
First, we have extracted text portions from the perspective documents of country
descriptions using NLTK based on occurrence of each search word that belong to a
decision factor. In order to get proper understanding of the context of the occur-
rence of the search word, we have extracted few sentences of text on both sides of
the occurrence location of search word. Altogether, we have extracted around
28,000 text portions from the total 21 perspectives documents. The accuracy of text
classifier purely depends on the quality of manually encoded training set data.
Therefore, in order to get a proper training set that is representative of the trans-
portation and logistics domain, we are planning to get 10 % of the extracted text
portions manually coded by the domain experts belong to Transportation and
Logistics domain. As shown in Fig. 11.1, major part of the manually coded texts
will be used for training the Naïve Bayes text classifier. After the classifier gets
sufficiently trained we will use the rest of the training set to test the accuracy of the
classifier. Finally, the classifier will be used to on the remaining 90 % of extracted
texts to classify them automatically using the machine learning technique. Fur-
thermore, for the text classification we would like to different domain-specific
models such as sentiment, emotion, and any other models that are specific and
suitable for the transportation and Logistics domain.
11.6 Illustrative Results
We present a selection of the results below to illustrate the methodological approach
and to demonstrate the domain-specific utility of applying text analytics methods
and techniques (Bird 2006; Yang and Liu 1999; Zhang and Li 2007). Table 11.1
presents the category analysis of the text corpus to extract the basic indicators of
global supply chain perspectives.
Figure 11.2 shows the five most frequent categories and their respective
cumulative keyword occurrences across the different time periods.
11 Methodological Demonstration of a Text Analytics Approach …125
Figure 11.3 shows the most frequent categories and measures for the whole data
corpus. Finally, Fig. 11.4 shows the emergent categories and measures for the
whole data corpus.
Table 11.1 Category analysis of global supply chain perspectives
Categories 2004–2007 2008–2009 2010–2013 2014
Waterways 1361 2232 2965 1117
Intermodal 668 701 1119 424
Economic structure 396 464 917 225
Railways 363 486 677 337
Economic policy 321 239 581 124
Airways 185 674 849 299
Geographical location 108 548 649 120
Fig. 11.2 Top 5 categories during time periods
Fig. 11.3 Top categories and measures for whole dataset
126 A. Kinra et al.
11.7 Discussion and Conclusion
Figure 11.5 presents the separation of concerns/division of labor between the
domain expert and the text analyst. As can be seen, the deep involvement of the
domain expert during the model specification and training phases of the text ana-
lytics process is necessary but not sufficient. Subsequent involvement of the domain
expert during the classification and model fine-tuning phases yields empirical
results that are robust, reliable, and relevant. We will further explore the dynamics
Fig. 11.4 Emergent
categories of sustainability
and security
Fig. 11.5 Separation of
concerns: domain expert
versus big data analyst
11 Methodological Demonstration of a Text Analytics Approach …127
of the relationship between specific domain expertise and generic big data analytics
expertise in our future work.
This study has developed and demonstrated an approach for automating the
information scanning, retrieval, and extraction process for logistics and trans-
portation applications. The approach is especially useful because it will aid in the
systematic study of big data while making sense of different types of logistics costs
that are currently being incurred in a tumultuous market place. Companies are
increasingly facing big data challenges in the global transportation of goods and
services, and look for dedicated approaches in this regard in order to leverage their
information system investments for optimizing their logistics and transportation
costs. For example, Maersk Line, the world’s largest container shipping company
has recently implemented a “Remote Container Management”system that collects
information using sensors and transceivers embedded in the containers. Such sys-
tems create a deluge of near real-time data that need to be modeled and analyzed
using methods and techniques from the emerging field of data science in general
and big data analytics in particular. However, we can also relate to the general
observation that model specification is the Achilles’Heel of big data analytics as it
is easy to under- or over-specify them without deep involvement of domain experts
and deep knowledge of domain-specific problems.
In addition to traditional text mining and machine learning methods, the next
steps involve employing the new method of “Social Set Analysis”to extract and
model latent social information and the resulting social influence, if any, from the
document corpus. Vatrapu et al. (2014a,b) have proposed a set theoretical approach
to big social data analytics called Social Set Analysis (SSA). SSA is based on the
sociology of associations and the mathematics of set theory and supports both
interaction analytics in terms of actors involved, actions taken, artifacts engaged
with as well as text analytics in terms of keywords employed, feelings expressed,
pronouns used, and topics discussed (Mukkamala et al. 2014a,b; Vatrapu et al.
2014a,b). For the purposes of this paper and the domains of trade and trans-
portation, we plan to employ SSA to uncover the temporal distribution of institu-
tional actors’engagement as well and unique actor/keyword sets before, during, and
after events of theoretical interest and the spatiotemporal dynamics of keywords and
expressed feelings with regard to country assessments in the text corpus.
References
Beugelsdijk S, Mudambi R, McCann P (2010) Place, space and organization: economic geography
and the multinational enterprise. J Econ Geogr 10(4):485–493
Beugelsdijk S, Mudambi R (2013) MNEs as border-crossing multi-location enterprises: The role
of discontinuities in geographic space. J Int Bus Stud 44
Bird S (2006) NLTK: the natural language toolkit. Paper presented at the proceedings of the
COLING/ACL on interactive presentation sessions
Bookbinder JH, Tan CS (2003) Comparison of Asian and European logistics systems. Int J Phys
Distribut Logist Manag 33(1)
128 A. Kinra et al.
Carter JR, Pearson JN, Peng L (1997) Logistics barriers to international operations: the case of the
people’s republic of China. J Bus Logist 18(2):129–145
Kinra A, Kotzab H (2008) Understanding and measuring macro-institutional complexity of
logistics systems environment. J Bus Logist 29(1):327–346
Kinra A (2015) Environmental complexity related information for the assessment of country
logistics environments: implications for spatial transaction costs and foreign location
attractiveness. J Transp Geogr 43:36–47
McCann P, Mudambi R (2004) The location behavior of the multinational enterprise: some
analytical issues. Growth Change 35(4):491–524
Memedovic O, Ojala L, Rodrigue JP, Naula T (2008) Fuelling the global value chains: what role
for logistics capabilities. Int J Technol Learn Innov Dev 1(3):353–374
Min H (1994) Location analysis of international consolidation terminals using the analytic
hierarchy process. J Bus Logist 15(2):25–44
Mukkamala R, Hussain A, Vatrapu R (2014a) Fuzzy-set based sentiment analysis of big social
data. In: Proceedings of IEEE 18th international enterprise distributed object computing
conference (EDOC 2014), Ulm, Germany, pp 71–80. doi:1510.1109/EDOC.2014.1519. ISBN:
1541-7719/1514
Mukkamala R, Hussain A, Vatrapu R (2014b) Towards a set theoretical approach to big data
analytics. In: Proceedings of the 3rd IEEE international congress on big data 2014, Anchorage,
United States
Rodrigue J-P (2012) The geography of global supply chains: evidence from third-party logistics.
J Supply Chain Manag 48(3):15–23
Rodrigue J-P, Notteboom T (2010) Comparative North American and European gateway logistics:
the regionalism of freight distribution. J Transp Geogr 18(4):497–507
Vatrapu R, Mukkamala R, Hussain A (2014a) A set theoretical approach to big social data
analytics: concepts, methods, tools, and findings. Paper presented at the computational social
science workshop at the European conference on complex systems 2014
Vatrapu R, Mukkamala R, Hussain A (2014b) Towards a set theoretical approach to big social data
analytics: concepts, methods, tools, and empirical findings. Paper presented at the 5th annual
social media & society international conference 2014
Yang Y, Liu X (1999) A re-examination of text categorization methods. Paper presented at the
proceedings of the 22nd annual international ACM SIGIR conference on research and
development in information retrieval
Zhang H, Li D (2007) Naïve Bayes text classifier. Paper presented at the granular computing,
2007. IEEE international conference on GRC 2007
11 Methodological Demonstration of a Text Analytics Approach …129
