Text Mining Methodologies with R: An

Application to Central Bank Texts

Jonathan Benchimol,† Sophia Kazinnik‡ and Yossi Saadon§

February 24, 2022

Abstract

We review several existing text analysis methodologies and explain their

formal application processes using the open-source software R and relevant
packages. Several text mining applications to analyze central bank texts are
presented.

Keywords: Text Mining, R Programming, Sentiment Analysis, Topic Modelling,

Natural Language Processing, Central Bank Communication, Bank of Israel.

JEL Codes: B40, C82, C87, D83, E58.

This paper does not necessarily reflect the views of the Bank of Israel, the Federal Reserve Bank
of Richmond or the Federal Reserve System. The present paper serves as the technical appendix of
our research paper (Benchimol et al., 2020). We thank Itamar Caspi, Shir Kamenetsky Yadan, Ariel
Mansura, Ben Schreiber, and Bar Weinstein for their productive comments.
† Bank of Israel, Jerusalem, Israel. Corresponding author. Email: jonathan.benchimol@boi.org.il
‡ Quantitative Supervision and Research, Federal Reserve Bank of Richmond, Charlotte, NC,

USA. Email: sophia.kazinnik@rich.frb.org

§ Research Department, Bank of Israel, Jerusalem, Israel. Email: yosis@boi.org.il

1
1 Introduction
The information age is characterized by the rapid growth of data, mostly unstruc-
tured data. Unstructured data is often text-heavy, including news articles, social
media posts, Twitter feeds, transcribed data from videos, as well as formal docu-
ments.1 The availability of this data presents new opportunities, as well as new
challenges, both to researchers and research institutions. In this paper, we review
several existing methodologies for analyzing texts and introduce a formal process
of applying text mining techniques using the open-source software R. In addition,
we discuss potential empirical applications.
This paper offers a primer on how to systematically extract quantitative infor-
mation from unstructured or semi-structured text data. Quantitative representa-
tion of text has been widely used in disciplines such as computational linguistics,
sociology, communication, political science, and information security. However,
there is a growing body of literature in economics that uses this approach to ana-
lyze macroeconomic issues, particularly central bank communication and financial
stability.2
The use of this type of text analysis is growing in popularity and has become
more widespread with the development of technical tools and packages facilitat-
ing information retrieval and analysis.3
An applied approach to text analysis can be described by several sequential
steps. Given the unstructured nature of text data, a consistent and repeatable ap-
proach is required to assign a set of meaningful quantitative measures to this type
of data. This process can be roughly divided into four steps: data selection, data
cleaning, information extraction, and analysis of that information. Our tutorial ex-
plains each step and shows how it can be executed and implemented using the
1 Usually in Adobe PDF or Microsoft Word formats.
2 See, for instance, Carley (1993), Ehrmann and Fratzscher (2007), Lucca and Trebbi (2009), Bholat

et al. (2015), Hansen and McMahon (2016), Bruno (2017), Bholat et al. (2019), Hansen et al. (2019),
Calomiris and Mamaysky (2020), Benchimol et al. (2021), Correa et al. (2021), and Ter Ellen et al.
(2022).
3 See, for instance, Lexalytics, IBM Watson AlchemyAPI, Provalis Research Text Analytics Soft-

ware, SAS Text Miner, Sysomos, Expert System, RapidMiner Text Mining Extension, Clarabridge,
Luminoso, Bitext, Etuma, Synapsify, Medallia, Abzooba, General Sentiment, Semantria, Kanjoya,
Twinword, VisualText, SIFT, Buzzlogix, Averbis, AYLIEN, Brainspace, OdinText, Loop Cognitive
Computing Appliance, ai-one, LingPipe, Megaputer, Taste Analytics, LinguaSys, muText, Tex-
tualETL, Ascribe, STATISTICA Text Miner, MeaningCloud, Oracle Endeca Information Discovery,
Basis Technology, Language Computer, NetOwl, DiscoverText, Angoos KnowledgeREADER, For-
est Rim’s Textual ETL, Pingar, IBM SPSS Text Analytics, OpenText, Smartlogic, Narrative Science
Quill, Google Cloud Natural Language API, TheySay, indico, Microsoft Azure Text Analytics API,
Datumbox, Relativity Analytics, Oracle Social Cloud, Thomson Reuters Open Calais, Verint Sys-
tems, Intellexer, Rocket Text Analytics, SAP HANA Text Analytics, AUTINDEX, Text2data, Saplo,
SYSTRAN, and many others.

2
open-source R software. For our sample data set, we use a set of monthly commu-
nications published by the Bank of Israel.
In general, an automatic and precise understanding of financial texts allows for
the construction of relevant financial indicators. There are many potential applica-
tions in economics and finance, as well as other social science disciplines. Central
bank publications (e.g., interest rate announcements, minutes, speeches, official re-
ports, etc.) are of particular interest, considering what a powerful tool central bank
communication is. This quick and automatic analysis of the underlying meaning
conveyed by these texts should allow for fine-tuning of these publications before
making them public. For instance, a spokesperson could use this tool to analyze
the orientation of a text, such as an interest rate announcement, before making it
public.
The remainder of the paper is organized as follows. The next section covers
theoretical background behind text analysis and interpretation of text. Section 3
describes text extraction and Section 4 presents methodologies for cleaning and
storing text data for text mining. Section 5 presents several common approaches
to text data structures used in Section 6, which details methodologies used for text
analysis, and Section 7 concludes.

2 Theoretical Background
The principal goal of text analysis is to capture and analyze all possible meanings
embedded in the text. This can be done both qualitatively and quantitatively. The
purpose of this paper is to offer an accessible tutorial to the quantitative approach.
In general, quantitative text analysis is a field of research that studies the ability to
decode data from natural language with computational tools.
Quantitative text analysis takes roots in a set of simple computational methods,
focused on quantifying the presence of certain keywords or concepts with a text.
These methods, however, fail to take into account the underlying meaning of text.
This is problematic because, as shown by Carley (1993), two identical sets of words
can have very different meanings. This realization and subsequent need to capture
meaning embedded in text gave rise to the development of new methods, such as
language network models, and, specifically, semantic networks (Danowski, 1993;
Diesner, 2013). Today, the common approach in quantitative text mining is to find
relationships between concepts, generating what is known as a semantic network.
Semantic network analysis is characterized by its ability to illustrate the rela-
tionships between words within a text, providing insights into its structure and

3
meaning. Semantic networks rely on co-occurrence metrics to represent proxim-
ity concepts (Diesner and Carley, 2011a,b; Diesner, 2013). For instance, nodes in a
network represent concepts or themes that frequently co-occur near each other in a
specific text. As a result, semantic network analysis allows meaning to be revealed
by considering the relationships among concepts.
In this paper, we cover both of the approaches mentioned above. We first
discuss term-counting methods, such as term frequency and relative frequency
calculations. We follow with networks-based methods, such as cluster analysis,
topic modeling, and latent semantic analysis. Overall, the field of natural lan-
guage processing (NLP) has progressed rapidly in recent years, but these methods
still remain to be essential and relevant building blocks of quantitative language
analysis.
The next three sections present a comprehensive set of steps for text analysis,
starting with common methodologies for cleaning and storing text, as well as dis-
cussing several common approaches to text data structures.

3 Text Extraction
For this exercise, we use a set of interest rate announcements published by the
Bank of Israel from 1997 to 2017. Overall, we have 220 documents of this type. We
use this set of documents as input using package tm (Feinerer et al., 2008) within
the open-source software R.4 This package can be thought as a framework for text
mining applications within R, including text preprocessing. There is a core func-
tion called Corpus embedded in the tm package. This function takes a predefined
directory, which contains the input (a set of documents) as an argument, and re-
turns the output, which is the set of documents organized in a particular way.
Here, we use the term corpus to reference a relevant set of documents.
We define our corpus in R in the following way. First, we apply a function
called file.path that defines a directory where all of our text documents are
stored.5 In our example, it is the folder that stores all 220 text documents, each cor-
responding to a separate interest rate decision meeting. After defining the working
directory, we apply the function Corpus from the package tm to all of the files in
the working directory. This way, the function captures and interprets each file as
a document and formats the set of text documents into a corpus object class as
4 Unnecessary elements (characters, images, advertisements, etc.) are removed from each docu-
ment to constitute our clean set of documents.
5 The folder should contain text documents only. If there are other files in that location (i.e., R

files) than the Corpus function will include the text in the other files.

4
defined internally by the tm package.

file.path <- file.path(".../data/folder")

corpus <- Corpus(DirSource(file.path))

Now, we have our documents organized as a corpus object. The content of each
document can be accessed and read using the writeLines function. For example:

writeLines(as.character(corpus[[1]]))

Using this command line, we can access and view the content of document
number one within the corpus. This document corresponds to the interest rate
discussion published in December 2007. Below are the first two sentences of this
document:

Bank of Israel Report to the public of the Bank of Israel’s discussions

before setting the interest rate for January 2007 The broad-forum
discussions took place on December 24, 2006, and the narrow forum
discussions on December 25, 2006, January 2007 General Before the Governor
makes the monthly interest rate decision, discussions are held at two
levels. The first discussion takes place in a broad forum, in which the
relevant background economic conditions are presented, including real and
monetary developments in Israel’s economy and developments in the global
economy.

There are other approaches to storing a set of texts in R, for example by using
the functions data.frame or tibble. However, we will concentrate on tm’s corpus
approach as it is largely more intuitive, and has a greater number of corresponding
functions explicitly written for text analysis.

4 Cleaning and Storing Text

Once the relevant corpus is defined, and read into the program, we can transform
it into an appropriate format for further analysis. As mentioned previously, each
document can be thought of as a set of tokens. Tokens are sets of words, numbers,
punctuation marks, and any other symbols present in the given document.

5
 Not all of the tokens carry meaningful information. Therefore, the next neces-
sary step is text cleaning, one of the crucial steps in text analysis. Text cleaning (or
text preprocessing) makes an unstructured set of texts uniform across and within
and eliminates idiosyncratic characters or meaningless terms.6 Text cleaning can
be loosely divided into a set of steps as shown below.
The text excerpt presented in Section 3 contains some useful information about
the content of the discussion, but also a lot of unnecessary details, such as punctu-
ation marks, dates, ubiquitous words. Therefore, the first logical step is to remove
punctuation and idiosyncratic characters from the set of texts. This includes any
strings of characters present in the text, such as punctuation marks, percentage or
currency signs, or any other characters that are not words.
One way to eliminate punctuation marks is with the removePunctuation func-
tion. This function removes a set of predefined punctuation characters.

corpus <- tm_map(corpus, removePunctuation)

The text below shows our original excerpt, with the aforementioned punctua-
tion characters removed:

The broad forum discussions took place on December 24 2006 and the narrow
forum discussions on December 25 2006 January 2007 General Before the
Governor makes the monthly interest rate decision discussions are held at
two levels The first discussion takes place in a broad forum in which the
relevant background economic conditions are presented including real and
monetary developments in Israel’s economy and developments in the global
economy

Additionally, any numbers present in the texts of our corpus can be removed
using the removeNumbers function, as in the below code:

corpus <- tm_map(corpus, removeNumbers)

Now, the text below shows our original excerpts, but without any punctuation
marks or digits:

6 Specific characters that are not used to understand the meaning of a text.

6
The broad forum discussions took place on December and the narrow forum
discussions on December January General Before the Governor makes the
monthly interest rate decision discussions are held at two levels The first
discussion takes place in a broad forum in which the relevant background
economic conditions are presented including real and monetary developments
in Israels economy and developments in the global economy

The current text excerpt is slightly more concise than the original, but there is
still much unnecessary information. Therefore, the next step would be to remove
the so-called stop words from the text.
What are stop words? Words such as “the”, “a”, “and”, “they”, and many
others can be defined as stop words. Stop words usually refer to the most common
words in a language, and because they are so common, they do not convey specific
information. Since these terms do not carry any meaning as standalone terms, they
are not valuable for our analysis. In addition to a pre-existing list of stop words,
ad hoc stop words can be added to the list.
We apply a function from the package tm onto our existing corpus as defined
above in order to remove the stop words. There is a coercing function called
removeWords that erases a given set of stop words from the corpus. There are
different lists of stop words available, and we use a standard list of English stop
words.
However, before removing the stop words, we need to turn all of our existing
words within the text into lowercase. Why? Because converting to lowercase, or
case folding, allows for case-insensitive comparison. This is the only way for the
function removeWords to identify the words subject for removal.
Therefore, using the package tm, and a coercing function tolower, we convert
our corpus to lowercase:

corpus <- tm_map(corpus, tolower)

Below is the example text excerpt following the command mentioned above:

the broad forum discussions took place on december and the narrow forum
discussions on december january general before the governor makes the
monthly interest rate decision discussions are held at two levels the first
discussion takes place in a broad forum in which the relevant background
economic conditions are presented including real and monetary developments
in israels economy and developments in the global economy

7
 We can now remove the stop words from the text:

corpus <- tm_map(corpus, removeWords, stopwords("english"))

Here, tm_map is a wrapper function that takes the corpus and applies character
processing function removeWords onto all of the contents of the corpus (all 220
documents). It returns the modified documents in the same format, a corpus, but
with the changes already applied. The following output shows our original text
excerpt with the stop words removed:

broad forum discussions took place december narrow forum discussions

december january general governor makes monthly interest rate decision
discussions held two levels first discussion takes place broad forum
relevant background economic conditions presented including real monetary
developments israels economy developments global economy

The next and final step is to stem the remaining words. Stemming is a process
of turning several words that mean the same thing into one. For example, after
stemming, words such as “banking”, “banks”, and “banked” become “bank”. As
stemming reduces inflected or derived words to their word stem or root. This can
be thought of as word normalization. Stemming algorithm allows us not to count
different variations as the same term as separate instances.
Below, we use a coercing function called stemDocument, that stems words in a
text document using Porter’s stemming algorithm (Porter, 1980). This algorithm
removes common morphological and inflectional endings from words in English,
as described in the previous paragraph.

corpus <- tm_map(corpus, stemDocument)

Once we have applied several of these character processing functions to our

corpus, we would like to examine it in order to view the results. Overall, as a
result of the above procedures, we end up with the following:

broad forum discuss took place decemb narrow forum discuss decemb januari
general governor make month interest rate decis discuss held two level
first discuss take place broad forum relev background econom condit present
includ real monetari develop israel economi develop global economi

8
 This last text excerpt shows what we end up with once the data cleaning ma-
nipulations are done. While the excerpt we end up with resembles its original only
remotely, we can still figure out reasonably well the subject of the discussion.

5 Data Structures
Once the text cleaning step is done, R allows for several convenient ways of storing
our results. We discuss the two most popular formats, a document term matrix and
a tidy object. While there may be more ways to store text, these two formats are
the most convenient when working with text data in R. We explain each of these
formats next.

5.1 Document Term Matrix

Document term matrix is a mathematical matrix that describes the frequency of
terms that occur in a collection of documents. Such matrices are widely used in
the field of NLP. In a document term matrix, each row corresponds to a specific
document in the collection and each column corresponds to the count of specific
terms within that document. In essence, a document term matrix is a vector-space
representation of a set of documents. An example of a document term matrix is
shown in Table 1.

Table 1. Document Term Matrix - Excerpt

Term j
Document i accord activ averag ...
May 2008 3 9 4 ...
June 2008 6 4 16 ...
July 2008 5 3 7 ...
August 2008 4 9 12 ...
September 2008 5 8 22 ...
October 2008 3 20 16 ...
November 2008 6 5 11 ...
... ... ... ... ...
Note: This table presents an excerpt of a document term matrix (dtm).

This type of matrix represents the frequency for each unique term in each doc-
ument in corpus. In R, our corpus can be mapped into a dtm object class by using
the function DocumentTermMatrix from the tm package.

9
dtm <- DocumentTermMatrix(corpus)

The goal of mapping the corpus onto a dtm is twofold; the first is to present the
topic of each document by the frequency of semantically significant and unique
terms, and second, to position the corpus for future data analysis. The value in
each cell of this matrix is typically the word frequency of each term in each doc-
ument. This frequency can be weighted in different ways. The default weighting
scheme function is called term frequency (tf).
Another common approach to weighting is called term frequency - inverse doc-
ument frequency (tf-idf). While the tf weighting scheme is defined as the num-
ber of times a word appears in the document, tf-idf is defined as the number of
times a word appears in the document but is offset by the frequency of the words
in the corpus, which helps to adjust for the fact that some words appear more
frequently in general.
Why is the frequency of each term in each document important? A simple
counting approach such as term frequency may be inappropriate because it can
overstate the importance of a small number of very frequent words. Term fre-
quency is the most normalized one, measuring how frequently a term occurs in a
document with respect to the document length, such as:

Number of times term t appears in a document

tf(t) = (1)
Total number of terms in the document
A more appropriate way to calculate word frequencies is to employ the tf-idf
weighting scheme. It is a way to weight the importance of terms in a document
based on how frequently they appear across multiple documents. If a term fre-
quently appears in a document, it is important, and it receives a high score. How-
ever, if a word appears in many documents, it is not a unique identifier, and it will
receive a low score. Equation 1 shows how words that frequently appear in a sin-
gle document will be scaled up, and Equation 2 shows how common words which
appear in many documents will be scaled down.

Total number of documents

idf(t) = ln (2)
Number of documents with term t in it

Conjugating these two properties yield the tf-idf weighting scheme:

tf-idf(t) = tf(t) idf(t) (3)

In order to employ this weighting scheme in Equation 3, we can assign this

10
option within the already familiar function dtm:

dtm <- DocumentTermMatrix(corpus, control = list(weighting =

weightTfIdf))

The above-mentioned steps provide us with a suitable numeric matrix under

the name of dtm. In this matrix, each cell contains an integer, corresponding to a
(perhaps weighted) number of times a specific term appeared in each document
within our corpus. However, most cells in such dtm will be empty, i.e., zeros,
because most terms do not appear in most of the documents. This property is
referred to as matrix sparsity.
Most term-document matrices tend to be high-dimensional and sparse. This is
because any given document will contain only a subset of unique terms that ap-
pear throughout the corpus. This will result in any corresponding document-row
having zeros for terms that were not used in that specific document. Therefore, we
need an approach to reduce dimensionality.
Function RemoveSparseTerms takes our existing dtm, and a certain numerical
value called sparse, and returns an adjusted dtm, where the number of elements is
reduced, depending on the value set for sparse. This numerical value is defined
as the maximum allowed sparsity, as is set in the range of zero to one.
In reference to RemoveSparseTerms, sparsity refers to the threshold of relative
document frequency of a term, above which the term will be removed. Sparsity
becomes smaller as this threshold approaches one. In a way, this process can be
thought of as removing outliers. For example, the code below will yield a dtm
where all terms appearing in at least 90% of the documents will be kept.

dtm <- RemoveSparseTerms(dtm, 0.1)

Now, we have a dtm that is ready for the initial text analysis. An example of
output following this weighting scheme and subsequent sparsity reduction of a
certain degree might yield Table 2.

5.2 Tidytext Table

This section provides an overview of the tidytext R package, detailed in Wickham
(2014). This framework was developed specifically for the R software, and for the
sole purpose of text mining. tidytext allows a set of documents to be presented

11
Table 2. Document Term Matrix - Excerpt with tf-idf

abroad acceler accompani account achiev adjust ...

01-2007 0.0002 0.000416 0.000844 0.000507 0.000271 0.000289 ...
01-2008 0.00042 0.000875 0.000887 0.000152 9.49E-05 0.000304 ...
01-2009 0.000497 0 0 9.01E-05 0.000112 0.000957 ...
01-2010 0.000396 0 0 7.18E-05 8.95E-05 0.000954 ...
01-2011 0.000655 0 0.000691 0.000119 7.39E-05 0.000552 ...
01-2012 0.000133 0 0.001124 9.65E-05 6.01E-05 0 ...
01-2013 0.00019 0.000395 0 0.000138 8.56E-05 0.000274 ...
01-2014 0 0.000414 0 0.000144 8.98E-05 0 ...
01-2015 0 0.00079 0 6.88E-05 8.57E-05 0.000183 ...
01-2016 0 0.000414 0 0 0.00018 0.000192 ...
01-2017 0 0.000372 0.000755 6.48E-05 0.000323 0.000689 ...
01-2018 0.000581 0 0.002455 0.000211 0 0 ...
Notes: This table presents an excerpt of a dtm with tf-idf weighting methodology. The highest
values for the selected sample are highlighted in gray.

as a one-term-per-document-per-row data frame. This is done with the help of the

unnest_tokens function within the tidytext.
This one-observation-per-row structure contrasts the way text is often stored
in alternative frameworks.7 For tidytext, the observation (or token) that is stored
in each row is most often a single term, but can also be an n-gram, sentence, or
paragraph. In general, a token refers to a unit by which we analyze the text.
Tokenization by word is one of the most intuitive units of analysis, but there
are cases where we may want to break the data into other units. For example,
an n-gram is an adjacent sequence of n terms from a given sample of text. One
might choose this unit when it is important to capture certain expressions, or sets
of words that go together.
For example, Figure 1 presents the most frequent words in the corpus as pro-
duced by the tidytext package. The below code selects words that appear in the
corpus at least 1200 times or more and plots their frequencies. Here, n is the word
count, i.e., how many times each word appears in the corpus:

tidy.table <- tidy.table %>%

count(word, sort = TRUE) %>%
filter(n > 1200) %>%
mutate(word = reorder(word, n)) %>%
7 Thisis different from other formats where each word corresponds with the document from
which it comes. In addition, it is possible to convert the tidytext format into the dtm format.

12
ggplot(aes(word, n)) + geom_col() + xlab(NULL) + coord_flip()

Figure 1. Histogram - tidytext

percent

rate

inflation

months

increased

growth

increase

bank

month

israel
0

2000

4000

6000

Note: Histogram containing most popular terms within the tidytext table.

Besides being more intuitive, the tidytext package has the capability for better
graphics.

5.3 Data Exploration

Given a dtm with reduced dimensions, as described above, we can apply exploratory
analysis techniques to find out what kind of information the corpus as a whole, or
each document within the corpus, contains. As with the text cleaning procedures,
there are several logical steps, and the first would be to find out what are the most
frequent terms within the dtm.
The following piece of code sums up the columns within the dtm, and then sorts
them in descending order within a data frame. We can then view terms with the six
highest and the lowest frequencies by using functions head and tail, respectively:

13
term.frequencies <- colSums(as.matrix(dtm))
order.frequencies <- order(term.frequencies)
head(order.frequencies, 6)
tail(order.frequencies, 6)

Table 3 is an example of an output following these commands, showing the top

six most frequent words and their corresponding frequencies.

Table 3. Top 6 Frequent Words in dtm

Term j
percen 9927
rate 9127
increas 5721
month 5132
interest 4861
bank 4039
Note: top six most frequent words with their corresponding frequencies.

Another, and perhaps easier, way to identify the frequency terms within the
dtm is to use the function findFreqTerms, which is a part of the package tm. This
function returns a list of terms, which meet two ad-hoc criteria of upper and lower
bounds of frequency limits:

findFreqTerms(dtm, lowfreq = 1800, highfreq = 5000)

Here is an example of an output following these commands. Figure 2 shows a

histogram of the terms that appear at least 1800 within our corpus.
Another way to look at the most popular terms is to use the dtm with the tf-idf
frequency weighted terms. Figure 3 shows a histogram of some of the most com-
mon terms within the corpus, as weighted by the tf-idf approach.
Once we know some of the most frequent terms in our corpus, we can explore
the corpus further by looking at different associations between some of them. This
can be done with the function findAssocs, part of the package tm. This func-
tion takes our dtm, and a user-selected specific term such as “bond”, “increas”,
“global”, etc., and an inclusive lower correlation limit as an input, and returns a
vector of matching terms and their correlations (satisfying the lower correlation
limit corlimit):

14
Figure 2. Histogram - Corpus

4000

3000
Frequencies

2000

1000

0
in ate

m as
th
co lat
u
in ar
ex st

ba t
de k
m lin

pr t
gr ice

is h
el
c

ke
n
in

t
re
pe

ra
on

ow
ye

c
e

f
nt

ar
in
r
cr

te

Corpus Terms

Note: Histogram containing most popular terms within the corpus.

findAssocs(dtm, "bond", corlimit = 0.5)

findAssocs(dtm, "econom", corlimit = 0.35)
findAssocs(dtm, "fed", corlimit = 0.35)

Table 4 is an example of an output following these for the term “bond”.

While this might not be the best way to explore the content of each text or the
corpus in general, it might provide some interesting insights for future analysis.
The math behind the findAssocs function is based on the standard function cor in
R’s stats package. Given two numeric vectors, corlimit computes their correla-
tion.
Another way to explore the contents of our corpus is to create a so-called word
cloud. A word cloud is an image composed of words used in a particular text or
corpus, in which the size of each word indicates its frequency. This can be done
with the use of the wordcloud package.8 Below, we plot word clouds using two dif-
ferent approaches to calculating the frequency of each term in the corpus. The first
8 https://cran.r-project.org/web/packages/wordcloud/wordcloud.pdf

15
Figure 3. Histogram - tf-idf

1.00

Frequencies w. tf−idf Weighting

0.75

0.50

0.25

0.00
a s
llio r
co ri n
m se
c r
ro lin
m ho se
pr neme
io ri
daus
ho ta
re c s
m pi
pain
s ut
di he st
pe ipl l
rc nin
en is
g
at cut
sen
t
sc ke
ms
bi rte

de pa
qu rea

ta

i
ev ta

ta
c
in

Corpus Terms

Note: Histogram containing most popular terms within the corpus, with tf-idf weighting.

approach uses a simple frequency calculation. We use the function set.seed here
to ensure all results and figures are reproducible. We can also vary the min.freq
parameter to select the minimum frequency threshold of each term.

set.seed(123)
wordcloud(names(freq), freq, min.freq = 400, colors=brewer.pal(8,
"Dark2"))

The function wordcloud provides a nice and intuitive visualization of the con-
tent of the corpus, or if needed, of each document separately. Figures 4 to 5 provide
several examples of an output following these commands. For instance, Figures 4
through 5 show word clouds containing word terms that appear at least 700, 1000,
and 2000, respectively.
Alternatively, to explore the frequency of terms within the corpus, one can use
the tf-idf weighting scheme, and reproduce the figures. It is clear that with the
new weighting scheme, other terms are emphasized more. As an example, Figures
6 through 7 show word clouds with word frequencies above 0.06, 0.08 and 0.1.

16
Table 4. Correlation - bond

Term j
yield 0.76
month 0.62
market 0.61
credit 0.60
aviv 0.58
rate 0.58
bank 0.57
indic 0.57
announc 0.56
sovereign 0.55
forecast 0.55
held 0.54
measur 0.54
treasuri 0.52
germani 0.52
index 0.51
Note: Terms most correlated with the term “bond”.

Figure 4. Word Cloud - 400 and 700 Terms

deficit exchangbackground
period global compar develop
lower econom
econom activ govern previous remain
financi economi
billion
global

averag stabil declin

lowbank
yield
yield
point expect target
last nis term level
continu

inflat quarter declin inflat indic monetari

one

bond
price index
capit
nomin
dollar
israel month tax target
billion
price
nis month growth
growth

rate rate bank

chang
expect

year
quarter

trend time cpi path rang

forecast
data

employ
first termlevel
economi

hous
increas
sinc

increas
develop

support israel
govern
polici

past real wage

year indic continu
market compar
high
point

home begin
deriv

interest market interest

activ
stabil
assess

basi
budget bond cpi
effect slight forecast
central rang index remain risk previous polici
season shekel monetari rise export
data
moder next percentag
sector

Note: Word cloud containing terms that appear at least 400 (left panel) and 700 (right panel) times
in the corpus, with word frequencies above 0.06.

17
Figure 5. Word Cloud - 1000 and 2000 Terms

growth
increasindic expect

stabil
continu bank month
month

rate
israel govern
inflat expect
economi
target
quarter
rate declin
price
term
continu
interest
year
inflat year index
interest bank point increas
market forecast
Note: Word cloud containing terms that appear at least 1000 and 2000 times in the corpus (left and
right panel, respectively), with word frequencies above 0.06.

Figure 6. Word Cloud - Word Frequencies Above 0.06

show money local term model

stabil worker law mid

export index percentag month

initi

affect past setorder ofebruari

move

context horizon
event

bond shekel
full

monetariact
estat
whole

aprilÃ
long data
explain

faster vat
seem

declin rise remain

fiscal

low
investor billion cpi
fuel

cut far
home

increas till
must

risk led lack

yet
hous

note
quarterfall
next

oil
howev
expect

short
indic aim fell
compar rose vis
eas

issuanc turn
aviv
just
reform fix

attain
mortgag
averag

previous nis rapid imflast valu

littl

tax
abil
disciplin global limitfed entir

deal

fact gap sustain

item earlier
season debt water

Note: Word cloud containing word terms with word frequencies above 0.06.

18
Figure 7. Word Cloud - Word Frequencies Above 0.08

non budget fix bond firstdebt upper

new money indic context
stock loan telset vis horizon
hous

long
order nevertheless previous fell
rapid rose declin line limit
serv oil till attain
note homequarter
past fall must act
crisi
occur

index
increas

show far
shekel cut

moder
short nis billion rise cpi

low
fiscal compar data eas
global monetari percentag step
averag remain disciplin led tax
local
model sustain commit ensur
half month explain season actual

express mortgag
season
framework percentag bondmoder
recoveri fiscal disciplin
debt model
sustain hous rose attain order
index data compar note
set
rapid
home quarter past
differenti

short nis
increas
monetari

limit
global

contextcpi
billionrise
remain declincut budget
must

export longprevious
shekel explainlocal
show averag indic
process money fall month
horizon ensur
circumst pressur central

Note: Word cloud containing word terms with word frequencies above 0.08 (top panel) and 0.1
(bottom panel).

19
 Another way to explore relationships between terms is to employ network
analysis, as it allows for a more comprehensive approach to exploring relation-
ships within the corpus terms. In general, network analysis allows texts to be
modeled as networks, with connections between nodes representing relations be-
tween them. A common way to model text in this framework is to apply a cluster-
ing algorithm to visualize it with a type of dendrogram or adjacency diagram. We
provide examples for both types next.
Clustering algorithms provide a way of identifying clusters of nodes that are
densely connected to each other on the network. The clustering method can be
thought of as an automatic text categorization. The basic idea behind document or
text clustering is to categorize documents into groups based on likeness criteria.
For example, one could calculate the Euclidean, or geometric, distance between
the terms to get a measure of likeness. Terms are then grouped on some distance
related criteria.
One of the most intuitive ways to visualize relationships between terms is with
correlation maps. Correlation maps show how some of the most frequent terms
relate to each other in the corpus, based on a certain ad-hoc correlation criteria.
While this is not a directed graph, like ones used in semantic networks, it still
reflects some information about the relationships between terms.
Below is the code example that can create a correlation map for a given dtm. To
plot this object, one will need to use the Rgraphviz package.9

correlation.limit <- 0.6

freqency.terms.dtm <- findFreqTerms(dtm.tf.idf)
plot(dtm.tf.idf, term = freqency.terms.dtm,
corThreshold=correlation.limit)

We plot Figures 8 and 9 following the above code.

Another way to visualize a network representing relationships between terms
within the corpus is to create a dendrogram. This technique arranges the net-
work into a hierarchy of groups according to a specified method. Hierarchical
clustering is an approach that allows to identifying groups in the dataset without
pre-specifying the number of clusters. Hierarchical clustering output can be visu-
alized in a tree-based representation of the observations, called a dendrogram. We
can plot a dendogram for our dtm using the following commands:

9 Available at: http://bioconductor.org/biocLite.R or through R console such as:

install.packages("BiocManager")
BiocManager::install("Rgraphviz")

20
Figure 8. Correlation Map Using dtm - Simple Counting Weighting Scheme

continu
declin bank
expect interest
growth israel
increas
market inflat
month
price
rate
year
Note: The vertical axis represents the heights.

dendogram <- dist(t(dtm.sparse.01), method = "euclidian")

dendogram.fit <- hclust(d = dendogram, method = "ward.D")
plot(dendogram.fit, hang = -1)

Here, the clustering method is based on the Ward’s method, which is based on
the original function approach. It relies on choosing a minimum variance criterion
that would minimize the total within-cluster variance (Ward, 1963).
The following dendograms presented in Figure 10 are built based on tf and
tf-idf weighting schemes, respectively.
Another quick and useful visualization of each document’s content is allowed
by heat maps. Heat maps can be used to compare the content of each document,
side by side, with other documents in the corpus, revealing interesting patterns
and time trends.
Figure 11 presents word frequencies for the word list on the bottom of the
heatmap. It demonstrates a simple distribution of word frequencies throughout
time. For example, the term accommod was used heavily during the discussions

21
that took place in mid and late 2001; however, it was not mentioned at all in early
2002.
Figure 12 presents a heatmap of word frequencies for the period spanning the
mid-1999 to early 2000. For example, the term “inflat”, representing discussion
around inflation, shows that this topic was discussed heavily in early 2000, in
particular in January of 2000. These kinds of figures provide a quick and visual
representation of any given interest rate discussion.
This section sums up some of the most popular techniques for exploratory text
analysis, demonstrating how a set of texts can be summarized and visualized in
an easy and intuitive way. In the next section, we talk about deriving text-based
quantitative metrics, allowing us to measure and summarize various aspects of
text.

6 Text Analytics
In this section we implement methods that allow us to conduct analysis both
within and between texts. Techniques such as sentiment analysis and term fre-
quency weighting can be used for the first purpose. The second group is used
for comparison between texts and refers to techniques related to Latent Seman-
tic Analysis (LSA) and Latent Dirichlet Allocation (LDA). We describe these tech-
niques in more detail in this section. The majority of text analytic algorithms in R
are written with the dtm format in mind. For this reason, we will use dtm format in
order to discuss the application of these algorithms.

6.1 Word Counting

Dictionary-based text analysis is a popular technique due to its simplicity. This
technique is based on selecting a predefined list of words that are relevant for the
analysis of that particular text. For example, the most commonly used source for
word classifications in the literature is the Harvard Psycho-sociological Dictionary,
specifically, the Harvard-IV-4 TagNeg (H4N) file. In general, dictionaries contain
lists of words that correspond to different categories, such as positive or negative
sentiment.
However, word categorization for one discipline (for example, psychology)
might not translate effectively into another discipline (for example, economics
or finance). Therefore, one of the drawbacks of this approach is the importance
of adequately choosing an appropriate dictionary or a set of predefined words.
Loughran and Mcdonald (2011) demonstrate that some words that may have a

22
negative connotation in one context may be neutral in others. The authors show
that dictionaries containing words like tax, cost, or liability that convey negative
sentiment in a general context, are more neutral in tone in the context of financial
markets. The authors construct an alternative, finance-specific dictionary to reflect
tone in a financial text better. They show that, with the use of a finance-specific
dictionary, they can predict asset returns better than other, generic, dictionaries.
In this tutorial, we use the Loughran and Mcdonald (2011) master dictionary,
which is available on their website. We divide the dictionary into two separate
csv files into two sentiment categories. Each file contains one column with several
thousand words; one is a list of positive terms, and one is a list of negative terms.
We read in both of these files into R as csv files.

dictionary.finance.negative <- read.csv("negative.csv",

stringsAsFactors = FALSE)[,1]
dictionary.finance.positive <- read.csv("positive.csv",
stringsAsFactors = FALSE)[,1]

For example, a word is a positive term if it belongs to a positive, or "hawkish"

category: for example, “increas”, “rais”, “tight”, “pressur”, “strength”, and so on.
A word is a negative term if it belongs to a negative or "dovish" category: for
example, “decreas”, “lower”, “loos”, “unsatisfi”, “worse”, and so on. A document
is classified as positive if the count of positive words is greater than or equal to
the count of negative words. Similarly, a document is classified as negative if the
count of negative words is greater than the count of positive words.
The constructed indicator is presented in Figure 14. We transform both files
into two separate data frames, using function data.frame. This function is used for
storing data tables in a matrix form. We apply the same text cleaning manipulation
to the dictionary words as applied to the corpus texts themselves. The code below
applies the text data cleaning principles to the two sentiment dictionaries that we
have uploaded. The cleaning involves turning all terms within both dictionaries
to lowercase, stemming all of the terms, and dropping all duplicate terms:

dictionary.negative <- tolower(dictionary.negative)

dictionary.negative <- stemDocument(dictionary.negative)
dictionary.negative <- unique(dictionary.negative)

23
 We then use the match function that compares the terms in both dictionaries
with each term in the corpus. This function returns a value of one if there is a
match, and a value of zero if there is no match. This allows us to calculate the
number of times each positive and each negative term appeared in the corpus. We
proceed to calculate the relative frequency of each dictionary terms. The code be-
low captures the list of terms from the dtm by using the function colnames and then
matches each of the terms in the corpus with the terms in each of the dictionaries,
calculating the total amount of matches for each dictionary:

corpus.terms <- colnames(dtm)

positive.matches <- match(corpus.terms, dictionary.positive,
nomatch=0)
negative.matches <- match(corpus.terms, dictionary.negative,
nomatch=0)

We then assign a value of one to each positive term (P) in the document, and
a value of minus one to each negative term (N) in a document, and measure the
overall sentiment score for each document i by the following formula:

Pi Ni
Scorei = 2 [ 1; 1] (4)
Pi + Ni

A document is classified as positive if the count of positive words is greater

than or equal to the count of negative words. Similarly, a document is negative if
the count of negative words is greater than the count of positive words. The code
below demonstrates a simple calculation of this indicator:

document.score = sum(positive.matches) - sum(negative.matches)

scores.data.frame = data.frame(scores = document.score)

Figure 13 presents the main indicators constructed using the dictionary word
count.

24
 Figure 9. Correlation Map Using dtm - tf-idf Frequency

committe countri cut drop inflationari month recoveri show upper crisi past imf mortgag primarili fight gas

member accommod fell econometr fall pressur debt

agre home transact model rise

25
vote scenario rose

project

Note: The vertical axis represents the heights.

Figure 10. Dendogram - tf and tf-idf Frequency
400
300

rate
200
Height

100

continu

year
growth

expect

inflat
israel

market

price
bank

interest
target
0

econom

govern
polici

stabil

dendogram
hclust (*, "ward.D")
0.015
0.010
Height

0.005

price
econom
stabil
year
expect
continu
growth
polici
0.000

market
govern

bank
inflat
target

Notes: Dendograms for tf-idf (top) and tf-idf (bottom) frequencies. The vertical axis represents
dendogram
the heights. hclust (*, "ward.D")

26
Figure 11. Heatmap tf-idf for 2010
Color Key
Frequencies
0 0.004
Value

2001−07−23

2001−06−25

2001−12−23

2001−11−26

2001−10−29

2002−02−25

2002−01−28

2002−05−27

2002−03−25

2001−09−24

2001−08−27

2002−04−22
an gh

nd

n
ar ci

ba a s
t
ac od
rd

ad tiv

a c

an or
ap ual

as nd

gr in

nk

be e

ild
th t
ac

al ffec

gi
n

s
e

ck tta
co

ch

se
ac

bu
ou

ou
va

ba
ba
m

ou
pr
n
m
co
ac

Note: Heatmap for documents published in 2010 (tf-idf weighted dtm). The color key corresponds
to probabilities for each topic being discussed during the corresponding interest rate decision meet-
ing.

27
Figure 12. Heatmap tf for 1999
Color Key

0 5 10
Value

1999−07−25

1999−10−25

1999−09−27

1999−06−28

1999−12−27

1999−11−22

1999−08−23

1998−08−06

2000−01−24

2000−02−21

2000−03−27

2000−04−24
gr ct
th

i
e
ex m

te
co nk
ec tinu

is t
m el

po t
in lat

lic
s

ke

ic
pe
ow
o

re
ra

ra
ba

pr
on

ar
in
n

te

Notes: Heatmap for documents published in 1999 (tf weighted dtm). The color key corresponds to
probabilities for each topic being discussed during the corresponding interest rate decision meet-
ing.

28
Figure 13. Sentiment - Word Count

0.009
Sentiment Score

0.006

0.003

0.000
2000 2005 2010 2015
Date

0.015

0.010
Sentiment Score

0.005

0.000

2000 2005 2010 2015

Date

0.0125

0.0100

0.0075
Sentiment Score

0.0050

0.0025

0.0000
2000 2005 2010 2015
Date

Notes: Scaled count of positive (top panel), negative (middle panel), and uncertainty (bottom panel)
words in each document using the dictionary approach.
29
 Using the positive and negative sentiment indicators exposed in Figure 13, Fig-
ure 14 shows the simple dictionary based sentiment indicator.

Figure 14. Sentiment Indicator - Dictionary Approach

0.01

0.00
Sentiment

−0.01

−0.02

2000 2005 2010 2015

Date

Note: Sentiment indicator built using the dictionary approach.

Figure 15 demonstrates a distribution of positive and negative matches through-

out the corpus, as produced by the package tidytext.
To sum it up, word counting is a relatively straightforward way to summa-
rize the sentiment of any given document. The strength of this approach is that
it is intuitive, and easy to implement. In addition, any given dictionary that is
being used for document scoring can be customized with ad-hoc words, related to
the subject matter. This, however, opens the door to a potential weakness of this
approach. There is a point where a customized dictionary list might lose its objec-
tivity. Dictionary-based sentiment measurement is the first step in the sentiment
extraction process, but there are more sophisticated ways to capture sentiment. We
discuss these in the next section.

6.2 Relative Frequency

In this section we discuss text mining techniques that take into account relative
word usage within documents to reveal information about text. Specifically, we
discuss a semi-supervised algorithm called wordscores, that estimates policy posi-
tions by comparing sets of texts using the underlying relative frequency of words.
This approach, described by Laver et al. (2003), proposes an alternative way to
locate the policy positions of political actors by analyzing the texts they generate.

30
Figure 15. Sentiment Contributions
constraining negative

limit declined
depends decline
bound deficit
required slowdown
entrench unemployment
commitment negative
entrenched downward
directive crisis
prevent weakened
imposed declines
positive uncertainty

stability risk
effective uncertainty
positive revised
improvement predictions
stable risks
strengthened probability
achieve volatility
encouragement apparently
strengthening variable
leading depends
0 250 500 750 0 250 500 750
Contribution to sentiment

Note: This figure represents how much each term contributes to the sentiment in each correspond-
ing category. These categories are defined as mutually exclusive. Constraining (top left), positive
(top right), negative (bottom left), and uncertainty (bottom right) sentiments are represented.

Mainly used in political sciences, it is a statistical technique for estimating the pol-
icy position based on word frequencies. The underlying idea is that relative word
usage within documents should reveal information of policy positions.
The algorithm assigns policy positions (or "scores") to documents on the basis
of word counts and known document scores (reference texts) via the computation
of "word scores". One assumption is that their corpus can be divided into two
sets (Laver et al., 2003). The first set of documents has a political position that
can be either estimated with confidence from independent sources or assumed
uncontroversial. This set of documents is referred to as the “reference” texts. The
second set of documents consists of texts with unknown policy positions. These
are referred to as the “virgin” texts. The only thing known about the virgin texts
is the words in them, which are then compared to the words observed in reference
texts with known policy positions.
One example of a reference text describes the interest rate discussion meeting
that took place on November 11, 2008. We chose this text as a reference because it is
a classic representation of dovish rhetoric. The excerpt below mentions a negative
economic outlook, both in Israel and globally, and talks about the impact of this
global slowdown on real activity in Israel:

Recently assessments have firmed that the reduction in global growth

31
will be more severe than originally expected. Thus, the IMF
significantly reduced its growth forecasts for 2009: it cut its
global growth forecast by 0.8 percentage points to 2.2 percent, and
its forecast of the increase in world trade by 2 percentage points, to
2.1 percent. These updates are in line with downward revisions by
other official and private-sector entities. The increased severity of
the global slowdown is expected to influence real activity in Israel.
The process of cuts in interest rates by central banks has intensified
since the previous interest rate decision on 27 October 2008.

Another example of the reference text describes the interest rate discussion
meeting that took place on June 24, 2002. This text is a classic representation of
hawkish rhetorics. For example, the excerpt below mentions a sharp increase in
inflation and inflation expectations:

The interest-rate hike was made necessary because, due to the rise in
actual inflation since the beginning of the year and the depreciation
of the NIS, inflation expectations for the next few years as derived
from the capital market, private forecasters, and the Bank of Israel’s
models have also risen beyond the 3 percent rate which constitutes the
upper limit of the range defined as price stability. Despite the two
increases in the Bank of Israel’s interest rate in June, inflation
expectations for one year ahead have risen recently and reached 5
percent.

Specifically, the authors use relative frequencies observed for each of the dif-
ferent words in each of the reference texts to calculate the probability that we are
reading a particular reference text, given that we are reading a particular word.
This makes it possible to generate a score of the expected policy position of any
text, given only the single word in question.
Scoring words in this way replaces and improves upon the predefined dictio-
nary approach. It gives words policy scores, without having to determine or con-
sider their meanings in advance. Instead, policy positions can be estimated by
treating words as data associated with a set of reference texts.10
10 However,one must consider the possibility that there would be a change in rhetoric over time.
Perhaps it would make sense to re-examine the approach at certain points in time. This would
depend on the time span of the data.

32
 In our analysis, out of the sample containing 220 interest rate statements, we
pick two reference texts that have a pronounced negative (or “dovish”) position
and two reference texts that have a pronounced positive (or “hawkish”) position
regarding the state of the economy during the corresponding month. We assign
the score of minus one to the two “dovish” reference texts and the score of one to
the two “hawkish” reference texts. We use these known scores to infer the score of
the virgin, or out of sample, texts. Terms contained by the out of sample texts are
compared with the words observed in reference texts, and then each out of sample
text is assigned a score, Wordscorei .
In R, we utilize the package quanteda, which contains the function wordfish.
This function takes a predefined corpus and applies the wordscores algorithm as
described above. Once the selection process of the reference documents is com-
plete, the code is fairly simple.

wordscore.estimation.results <- wordfish(corpus, dir = c(1,5))

This function takes our corpus as an input, as well as the two selected reference
documents (here, document number one and document number five), and returns
a set of estimation position, as related to each document.

6.3 Latent Semantic Analysis

Latent semantic analysis (LSA) is an unsupervised machine learning algorithm
that allows for mapping texts into a vector space (Deerwester et al., 1990). LSA
evaluates documents in order to find underlying meaning or concept of these doc-
uments. LSA can be used for many things, such as comparing document similarity
or clustering documents based on an ad-hoc criterion. LSA can decompose a doc-
ument term matrix into a reduced vector space that is assumed to reflect semantic
structure.
LSA is similar to approaches described in the previous section, but one of its
advantages is that it is fully unsupervised, and therefore does not rely on any sub-
jective elements. In R, one can implement LSA with the use of quanteda package,
thus creating a representation of documents in the reduced 2-dimensional space:

lsa_model <- textmodel_lsa(dtm)

lsa_predict <- predict(lsa_model)

33
 The output of an estimated LSA model allows for easier comparison between
documents, as well as mapping capabilities. Overall, there are various down-
stream tasks that can be implemented once the latent semantic space is created.

6.4 Topic Models

Topic modeling is a method for classifying and mapping text into a given number
of topics. Largely, topic modeling is a set of algorithms that automatically find and
classify recurring patterns of words throughout a set of documents. In this paper,
we use the LDA algorithm for the above task, which is a widely used algorithm in
the field of computer science (Blei et al., 2003).
This algorithm assumes that each of the documents within our corpus consists
of a mixture of corpus-wide topics. These topics, however, are not observable but
are instead hidden behind everyday words and sentences. Specifically, LDA esti-
mates what fraction of each document in a corpus can be assigned to each of the
topics. The number of topics is set in advance. We do not observe the topics in
the document, only the words that those topics tend to generate. LDA is an al-
gorithm that employs an unsupervised learning approach, in that we do not set
prior probabilities for any of the words belonging to any given topic. Besides, it
is a mixed-membership model, and therefore the assumption is that every word
in the corpus simultaneously belongs to several topics and the topic distributions
vary over documents in the corpus.
To provide more intuition, consider an implicit assumption that a given set of
words relates to a specific topic. For example, consider the following set of words:
gain, employment, labor. Each of these words would map into an underlying topic
“labor market” with a higher probability compared to what it would map into the
topic of “economic growth”. This algorithm has a considerable advantage, its ob-
jectivity. It makes it possible to find the best association between words and the
underlying topics without preset word lists or labels. The LDA algorithm works
its way up through the corpus. It first associates each word in the vocabulary to
any given latent topic. It allows each word to have associations with multiple top-
ics. Given these associations, it then proceeds to associate each document with
topics. Besides the actual corpus, the main input that the model receives is how
many topics there should be. Given those, the model will generate βk topic distrib-
utions, the distribution over words for each topic. The model will also generate θ d
document distributions for each topic, where d is the number of documents. This
modeling is done using Gibbs sampling iterations, going over each term in each
document and assigning relative importance to each instance of the term.

34
 In R, we use the package topicmodels, with the default parameter values sup-
plied by the LDA function. Specifying a parameter is required before running the
algorithm, which increases the subjectivity level. This parameter, k, is the number
of topics that the algorithm should use to classify a given set of documents. There
are analytical approaches to decide on the values of k, but most of the literature set
it on an ad hoc basis. When choosing k we have two goals that are in direct conflict
with each other. We want to correctly predict the text and be as specific as possible
to determine the number of topics. Yet, at the same time, we want to be able to
interpret our results, and when we get too specific, the general meaning of each
topic will be lost.
Let us demonstrate with this example by first setting k = 2, meaning that we
assume only two topics to be present throughout our interest rate discussions.
Below are the top seven words to be associated with these two topics. It can be
seen below that while these two sets of words differ, they both have overlapping
terms. This demonstrates that each word can be assigned to multiple topics but
with a different probability.

Table 5. Topic-relevant Words - Topics 1 to 2

Topic 1 Topic 2
"increas" "rate"
"rate" "interest"
"month" "expect"
"continu" "israel"
"declin" "inflat"
"discuss" "bank"
"market" "month"
"monetari" "quarter"
Note: Terms with the highest probability of appearing in Topic 1 and Topic 2.

Table 5 shows that Topic 1 relates directly and clearly to changes in the tar-
get rate, while Topic 2 relates more to inflationary expectations. However, these
are not the only two things that the policymakers discuss during interest rate meet-
ings, and we can safely assume that there should be more topics considered, mean-
ing k should be larger than two.11
To demonstrate the opposite side of the trade-off, let us consider k = 6, i.e., we
assume six different topics are being discussed. Below is the top seven words with
the highest probability to be associated with these six topics:
11 The supervised approach may help to determine the main theme of each topic objectively.

35
Table 6. Topic-relevant Words - Topics 1 to 6

Topic 1 Topic 2 Topic 3 Topic 4 Topic 5 Topic 6

"declin" "bank" "increas" "continu" "quarter" "interest"
"monetari" "economi" "month" "rate" "year" "rate"
"discuss" "month" "interest" "remain" "rate" "israel"
"rate" "forecast" "inflat" "market" "growth" "inflat"
"data" "market" "hous" "term" "month" "expect"
"polici" "govern" "continu" "year" "expect" "discuss"
"indic" "global" "rate" "price" "first" "bank"
"develop" "activ" "indic" "growth" "point" "econom"
Note: Terms with the highest probability of appearing in Topics 1 through 6.

The division between topics is less clear in Table 6 compared to Table 5. While
Topics 1, 2 and 3 relate to potential changes in interest rate, Topic 4 relates to hous-
ing market conditions, and Topic 5 relates to a higher level of expected growth tak-
ing into account monetary policy considerations. Topic 6 covers economic growth
and banking discussions. We see that while we get more granularity in topics
by increasing the possible number of topics, we see increased redundancy in the
number of topics. Given this outcome, we could continue to adjust k and assess
the result.
Next, we demonstrate how we run this algorithm. First, we specify a set of
parameters for Gibbs sampling. These include burnin, iter, thin, which are the
parameters related to the amount of Gibbs sampling draws, and the way these are
drawn.

burnin <- 4000

iter <- 2000
thin <- 500
seed <- list(2003, 5, 63, 100001, 765)
nstart <- 5

As discussed, the number of topics is decided arbitrarily, as an educated guess,

and can be adjusted as needed. Here, based on the previous analysis we take the
average of the two previous assumptions and end up with four assumed topics.

k <- 4

36
 The code below runs the LDA algorithm, using the set of parameters as de-
scribed above and our dtm:

lda.results <- LDA(dtm.sparse, k, method = "Gibbs", control =

list(nstart = nstart, seed = seed, best = best, burnin = burnin, iter
= iter, thin = thin))

These last lines write out and save the estimated topics as provided by the LDA
algorithm:

lda.topics <- as.matrix(topics(lda.results))

lda.results.terms <- as.matrix(terms(lda.results,8))
lda.results.terms

Table 7. Topic-relevant Words - Topics 1 to 4

Topic 1 Topic 2 Topic 3 Topic 4

"expect" "increas" "interest" "month"
"continu" "declin" "rate" "increas"
"rate" "continu" "stabil" "rate"
"inflat" "rate" "israel" "forecast"
"interest" "expect" "bank" "bank"
"rang" "remain" "inflat" "indic"
"israel" "growth" "market" "growth"
"last" "term" "govern" "year"
"price" "nis" "year" "previous"
"bank" "year" "target" "index"
"econom" "data" "term" "hous"
Note: Words with the highest probability of appearing in Topics 1 through 4.

Let us examine the topics presented in Table 7. Topic 1 relates to current changes
in interest rate and its goal of keeping inflation in range. It mentions the interest
rate, inflation expectations and the range of inflation. Topic 2 relates to the actual
inflation data and inflationary expectations. Topic 3 relates to a high-level mon-
etary policy discussion, and Topic 4 relates to housing market conditions. LDA
algorithm also generates probability distribution of topics over the corpus.
Figure 16 is a heat map containing a sample for the period of June 2007 to
April 2008. Given an assumption that only four topics are discussed during each

37
interest rate meeting, the values presented in the legend are probabilities for each
topic being discussed during the corresponding interest rate decision meeting.
For example, during the meeting of November 2008, the "Monetary Policy"
topic was discussed with greater probability compared to the "Inflation" topic. As
can be seen from Figure 16, this occurrence stands out from the regular pattern.
Figures 17 and 18 present the heat maps about the interest rate announcements
during the year 2007 and 2000, respectively.
Figure 17 shows that in a given set of documents, the bulk of the discussion
was spent on discussing the key interest rate set by the Bank of Israel. In contrast,
it can be seen that inflation was not discussed at all during certain periods. Figure
18 shows that the subject of discussion was mainly monetary policy during this
period of time.
The remaining question is how to determine whether the topics selected by
the LDA algorithm are a good fit for the underlying text. For that, we can use
topic coherence. Topic coherence allows for a set of words, as generated by a topic
model, to be rated for coherence or interpretability, as shown in Newman et al.
(2010). One way to implement this in R is to use the topicdoc package, based on
the paper by Boyd-Graber et al. (2014). This package can be applied on top of the
analysis described in the previous section, and contains many helpful functions
that allow for diagnostics of the estimated LDA model.
One can run a full set of diagnostics using the topic_diagnostics function.
This function takes the fitted LDA model as an input, and produces a set of mea-
sures that evaluate the fit, such as mean token length, exclusivity, coherence, and
document prominence of each topic.
Another way to explore and evaluate the output of an LDA model is to visual-
ize it. In R, this can be done with the help of LDAvis package (Sievert and Shirley,
2014; Mabey, 2018). LDA visualization can help answer questions such as what
is the meaning of each topic; how prevalent is each topic; and how the estimated
topics relate to each other. Together with the shiny package, the LDAvis provides
an interactive tool that can be used to create visualizations such as the intertopic
distance map.
The intertopic distance map is a visualization of the topics in a two-dimensional
space. The area of these topic circles is proportional to the number of words be-
longing to each topic across the corpus. The circles are plotted using a multidi-
mensional scaling algorithm based on the words they comprise, so topics that are
closer together have more words in common. This algorithm also considers the
saliency of terms, identifying the most informative or useful words for identify-
ing topics within the corpus. Higher saliency values indicate that a word is more

38
helpful in identifying a specific topic than a randomly selected term.

7 Conclusion
In this paper, we review some of the most commonly used text mining methodolo-
gies. We demonstrate how text sentiment and topics can be extracted from a set of
text sources. Taking advantage of the open-source software package R and related
packages, we provide a detailed step-by-step tutorial, including code excerpts that
are easy to implement and examples of output. The framework we demonstrate
in this paper shows how to process and utilize text data in an objective and auto-
mated way.
As described, the ultimate goal of this tutorial is to show how text analysis can
help uncover the information embedded in monetary policy communication and
to be able to organize it consistently. We first show how to set up a directory and
input a set of relevant files into R. We show how to represent and store this set of
files as a corpus, allowing for easy text manipulations. We then describe a series of
text cleaning procedures that sets the stage for further text analysis. In the second
part of the paper, we demonstrate approaches to initial text analysis, showing how
to create several summary statistics for our existing corpus. We then describe two
common approaches to text sentiment extraction and one common approach to
topic modeling.
We also consider term-weighting and contiguous sequence of words (n-grams)
to better capture the subtlety of central bank communication. We consider field-
specific weighted lexicons, consisting of two, three, or four-word clusters, relating
to a specific policy term being discussed. This approach provides an alternative
and sometimes more precise picture of the text content, as opposed to individual
terms, and allows us to find underlying patterns and linkages within text more
precisely.
NLP methods have come a long way in the past few decades. These methods
continue to evolve rapidly, getting increasingly more adept at capturing and un-
derstanding the nuances and context of language. At the same time, the ability
to capture signals such as sentiment, topics, and semantic relationships in the text
remains an important and relevant building block of language processing.

39
Figure 16. Probability Distribution - Topics 1 to 4 from 2007 to 2008

Color Key

0.1 0.3 0.5

Value

2009−06−22
2009−03−23
2008−09−22
2008−10−27
2009−08−24
2009−05−25
2009−07−27
2008−08−25
2008−11−11
2009−02−23
2009−01−26
2009−04−27
2008−11−24
2008−12−29
t
e

y
n

ke
lic
at

tio

ar
R

Po
fla

M
y

In
Ke

ng
ar
et

si
on

ou
M

Note: The color key corresponds to probabilities for each topic being discussed during the corre-
sponding interest rate decision meeting.

40
Figure 17. Probability Distribution - Topics 1 to 4 from 2008 to 2009

Color Key

0.1 0.4
Value

2008−02−25
2008−01−28
2008−05−26
2008−07−28
2007−10−29
2008−03−24
2007−12−24
2007−09−24
2008−06−23
2007−11−26
2008−04−28
2007−08−27
2008−08−25
t
e

y
n

ke
lic
at

tio

ar
R

Po
fla

M
y

In
Ke

ng
ar
et

si
on

ou
M

Note: The color key corresponds to probabilities for each topic being discussed during the corre-
sponding interest rate decision meeting.

41
Figure 18. Probability Distribution - Topics 1 through 4 from 1999 to 2000

Color Key

0.1 0.4
Value

2000−01−24

1999−10−25

1999−07−25

2000−04−24

2000−03−27

1999−12−27

2000−02−21

1999−11−22

1999−06−28

1999−09−27

1998−08−06

1999−08−23
t
e

y
n

ke
lic
at

tio

ar
R

Po
fla

M
y

In
Ke

ng
ar
et

si
on

ou
M

Note: The color key corresponds to probabilities for each topic being discussed during the corre-
sponding interest rate decision meeting.

42
References
Benchimol, J., Kazinnik, S., and Saadon, Y. (2020). Communication and trans-
parency through central bank texts. Paper presented at the 132nd Annual Meet-
ing of the American Economic Association, January 3-5, 2020, San Diego, CA,
United States.

Benchimol, J., Kazinnik, S., and Saadon, Y. (2021). Federal Reserve Communication
and the COVID-19 Pandemic. Covid Economics, 79:218–256.

Bholat, D., Broughton, N., Ter Meer, J., and Walczak, E. (2019). Enhancing central
bank communications using simple and relatable information. Journal of Mone-
tary Economics, 108(C):1–15.

Bholat, D., Hans, S., Santos, P., and Schonhardt-Bailey, C. (2015). Text mining for
central banks. Number 33 in Handbooks. Centre for Central Banking Studies,
Bank of England.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent Dirichlet allocation. Journal
of Machine Learning Research, 3:993–1022.

Boyd-Graber, J., Mimno, D., and Newman, D. (2014). Care and feeding of topic
models: Problems, diagnostics, and improvements. In Airoldi, E. M., Blei, D.,
Erosheva, E. A., and Fienberg, S. E., editors, Handbook of Mixed Membership Mod-
els and Their Applications, CRC Handbooks of Modern Statistical Methods, pages
225–254. CRC Press, Boca Raton, Florida.

Bruno, G. (2017). Central bank communications: information extraction and se-

mantic analysis. In Bank for International Settlements, editor, Big Data, vol-
ume 44 of IFC Bulletins chapters, pages 1–19. Bank for International Settlements.

Calomiris, C. W. and Mamaysky, H. (2020). How natural language processing will

improve central bank accountability and policy. Cato Journal, 40(2):447–465.

Carley, K. (1993). Coding choices for textual analysis: A comparison of content

analysis and map analysis. Sociological Methodology, 23:75–126.

Correa, R., Garud, K., Londono, J. M., and Mislang, N. (2021). Sentiment in central
banks’ financial stability reports. Review of Finance, 25(1):85–120.

Danowski, J. A. (1993). Network analysis of message content. Progress in Commu-

nication Sciences, 12:198–221.

43
Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer, T. K., and Harshman, R.
(1990). Indexing by latent semantic analysis. Journal of the American Society for
Information Science, 41(6):391–407.

Diesner, J. (2013). From texts to networks: detecting and managing the impact of
methodological choices for extracting network data from text data. KI - Kün-
stliche Intelligenz, 27(1):75–78.

Diesner, J. and Carley, K. M. (2011a). Semantic Networks. In Golson, G. B. . J., ed-

itor, Encyclopedia of Social Networking, pages 766–769. Thousand Oaks, CA: Sage
Publications.

Diesner, J. and Carley, K. M. (2011b). Words and Networks. In Golson, G. B.

. J., editor, Encyclopedia of Social Networking, pages 958–961. Thousand Oaks, CA:
Sage Publications.

Ehrmann, M. and Fratzscher, M. (2007). Communication by central bank commit-

tee members: Different strategies, same effectiveness? Journal of Money, Credit
and Banking, 39(2-3):509–541.

Feinerer, I., Hornik, K., and Meyer, D. (2008). Text mining infrastructure in R.
Journal of Statistical Software, 25(i05).

Hansen, S. and McMahon, M. (2016). Shocking language: Understanding the

macroeconomic effects of central bank communication. Journal of International
Economics, 99(S1):114–133.

Hansen, S., McMahon, M., and Tong, M. (2019). The long-run information effect of
central bank communication. Journal of Monetary Economics, 108(C):185–202.

Laver, M., Benoit, K., and Garry, J. (2003). Extracting policy positions from political
texts using words as data. American Political Science Review, 97(02):311–331.

Loughran, T. and Mcdonald, B. (2011). When is a liability not a liability? Textual

analysis, dictionaries, and 10-Ks. Journal of Finance, 66(1):35–65.

Lucca, D. O. and Trebbi, F. (2009). Measuring central bank communication: an au-

tomated approach with application to FOMC statements. NBER Working Papers
15367, National Bureau of Economic Research.

Mabey, B. (2018). pyLDAvis: Python Library for Interactive Topic Model Visual-
ization. Port of the R package LDAvis.

44
Newman, D., Lau, J. H., Grieser, K., and Baldwin, T. (2010). Automatic evaluation
of topic coherence. In Human Language Technologies: The 2010 Annual Conference
of the North American Chapter of the Association for Computational Linguistics, HLT
’10, page 100âĂŞ108. Association for Computational Linguistics.

Porter, M. F. (1980). An algorithm for suffix stripping. Program, 14(3):130–137.

Sievert, C. and Shirley, K. (2014). LDAvis: a method for visualizing and inter-
preting topics. In Proceedings of the Workshop on Interactive Language Learning,
Visualization, and Interfaces, pages 63–70, Baltimore, Maryland, USA. Association
for Computational Linguistics.

Ter Ellen, S., Larsen, V. H., and Thorsrud, L. A. (2022). Narrative monetary policy
surprises and the media. Journal of Money, Credit and Banking, Forthcoming.

Ward, J. H. (1963). Hierarchical grouping to optimize an objective function. Journal

of the American Statistical Association, 58(301):236–244.

Wickham, H. (2014). Tidy data. Journal of Statistical Software, 59(i10):1–23.

45