Learning objectives

Datacamp

• Sentiment analysis in R the Tidy way
  • Just the first chapter is required
  • You are welcome to do more, of course
• I will generally follow the same “tidy text” principles as the Datacamp course does – the structure keeps things easy to manage
  • We will sometimes deviate to make use of certain libraries, which, while less tidy, make our work easy than the corresponding tidy-oriented packages (if they even exist!)

Notes on the homework

• A few clarifications based on your emails:
  1. Exercise 1: The distribution of class action lawsuits by year only need to show the year and the number of lawsuits that year
  2. Exercise 2: The percent of firm-year observations with lawsuits b industry should have 4 calculations:
     • Ex.: (# of retail lawsuits) / (# of retail firm years)
  3. Exercise 3: The coefficient to explain is the coefficent of legal on fps – the only coefficient in the model

Review of Session 7

• Last session we saw that textual measures can help improve our fraud detection algorithm
• We looked at a bunch of textual measures:
  • Sentiment
  • Readability
  • Topic/content
• We didn’t see how to make these though…
  • Instead, we had a nice premade dataset with everything already done

We’ll get started on these today – sentiment and readability

We will cover making topic models in a later session

Why is textual analysis harder?

• Thus far, everything we’ve worked with is what is known as structured data
  • Structured data is numeric, nicely indexed, and easy to use
• Text data is unstructured
  • If we get an annual report with 200 pages of text…
    • Where is the information we want?
    • What do we want?
    • How do we crunch 200 pages into something that is…
      1. Manageable?
      2. Meaningful?

This is what we will work on today, and we will revist some of this in the remaining class sessions

Structured data

• Our long or wide format data

## # A tibble: 3 x 3
##   quarter level_3         value
##   <chr>   <chr>           <chr>
## 1 1995-Q1 Wholesale Trade 17   
## 2 1995-Q1 Retail Trade    -18  
## 3 1995-Q1 Accommodation   16

## # A tibble: 3 x 4
##   RegionID `1996-04` `1996-05` `1996-06`
##      <int>     <int>     <int>     <int>
## 1    84654    334200    335400    336500
## 2    90668    235700    236900    236700
## 3    91982    210400    212200    212200

The structure is given by the IDs, dates, and variables

Unstructured data

• Text
  • Open responses to question, reports, etc.
  • What it isn’t:
    • "JANUARY", "ONE", "FEMALE"
    • Months, numbers, genders
    • Anything with clear and concise categories
• Images
  • Satellite imagery
• Audio
  • Phone call recordings
• Video
  • Security camera footage

All of these require us to determine and impose structure

Some ideas of what we can do

1. Text extraction
   • Find all references to the CEO
   • Find if the company talked about global warming
   • Pull all telephone numbers or emails from a document
2. Text characteristics
   • How varied is the vocabulary?
   • Is it positive or negative (sentiment)
   • Is it written in a strong manner?
3. Text summarization or meaning
   • What is the content of the document?
   • What is the most important content of the document?
   • What other documents discuss similar issues?

Where might we encounter text data in business

1. Business contracts
2. Legal documents
3. Any paperwork
4. News
5. Customer reviews or feedback
   • Including transcription (call centers)
6. Consumer social media posts
7. Chatbots and AI assistants

Natural Language Processing (NLP)

• NLP is the subfield of computer science focused on analyzing large amounts of unstructured textual information
  • Much of the work builds from computer science, linguistics, and statistics
• Unstructured text actually has some structure – language
  • Word selection
  • Grammar
  • Word relations
• NLP utilizes this implicit structure to better understand textual data

NLP in everyday life

• Autocomplete of the next word in phone keyboards
  • Demo below from Google’s blog
• Voice assistants like Google Assistant, Siri, Cortana, and Alexa
• Article suggestions on websites
• Search engine queries
• Email features like missing attachment detection

Case: How leveraging NLP helps call centers

• How Analytics, Big Data and AI Are Changing Call Centers Forever
• Short link: rmc.link/420class8

What are call centers using NLP for?

How does NLP help call centers with their business?

Consider

Where an we make use of NLP in business?

• We can use it for call centers
• We can make products out of it (like Google and other tech firms)
• Where else?

Before we begin: Special characters

• Some characters in R have special meanings for string functions
  • \ | ( ) [ { } ^ $ * + ? . !
• To type a special character, we need to precede it with a \
  • Since \ is a special character, we’ll need to put \ before \…
    • To type $, we would use \\$
• Also, some spacing characters have special symbols:
  • \t is tab
  • \r is newline (files from Macs)
  • \r\n is newline (files from Windows)
  • \n is newline (files from Unix, Linux, etc.)

Loading in text data from files

• Use read_file() from tidyverse’s readr package to read in text data
• We’ll use Citigroup’s annual report from 2014
  • Note that there is a full text link at the bottom which is a .txt file
  • I will instead use a cleaner version derived from the linked file
    • The cleaner version can be made using the same techniques we will discuss today

# Read text from a .txt file using read_file()
doc <- read_file("../../Data/0001104659-14-015152.txt")
# str_wrap is from stringr from tidyverse
cat(str_wrap(substring(doc,1,500), 80))

## UNITED STATES SECURITIES AND EXCHANGE COMMISSION WASHINGTON, D.C. 20549 FORM
## 10-K ANNUAL REPORT PURSUANT TO SECTION 13 OR 15(d) OF THE SECURITIES EXCHANGE
## ACT OF 1934 For the fiscal year ended December 31, 2013 Commission file number
## 1-9924 Citigroup Inc. (Exact name of registrant as specified in its charter)
## Securities registered pursuant to Section 12(b) of the Act: See Exhibit 99.01
## Securities registered pursuant to Section 12(g) of the Act: none Indicate by
## check mark if the registrant is a

Loading from other file types

• Ideally you have a .txt file already – such files are generally just the text of the documents
• Other common file types:
  • HTML files (particularly common from web data)
    • You can load it as a text file – just note that there are html tags embedded in it
      • Things like <a>, <table>, <img>, etc.
    • You can load from a URL using RCurl
    • In R, you can use XML or rvest to parse out specific pieces of html files
    • If you use python, use lxml or BeautifulSoup 4 (bs4) to quickly turn these into structured documents

Loading from other file types

• Ideally you have a .txt file already – such files are generally just the text of the documents
• Other common file types:
  • PDF files
    • Use pdftools and you can extract text into a vector of pages of text
    • Use tabulizer and you can extract tables straight from PDF files!
      • This is very painful to code by hand without this package
      • The package itself is a bit difficult to install, requiring Java and rJava, though

Example using html

library(RCurl)
library(XML)

html <- getURL('https://coinmarketcap.com/currencies/ethereum/')
cat(str_wrap(substring(html, 46320, 46427), 80))

## n class="h2 text-semi-bold details-panel-item--price__value" data-currency-
## value>208.90</span> <span class="

xpath <- '//*[@id="quote_price"]/span[1]/text()'
hdoc = htmlParse(html, asText=TRUE)  # from XML
price <- xpathSApply(hdoc, xpath, xmlValue)
print(paste0("Ethereum was priced at $", price,
             " when these slides were compiled"))

## [1] "Ethereum was priced at $208.90 when these slides were compiled"

Automating crypto pricing in a document

# The actual version I use (with caching to avoid repeated lookups) is in the appendix
cryptoMC <- function(name) {
  html <- getURL(paste('https://coinmarketcap.com/currencies/',name,'/',sep=''))
  xpath <- '//*[@id="quote_price"]/span[1]/text()'
  hdoc = htmlParse(html, asText=TRUE)
  plain.text <- xpathSApply(hdoc, xpath, xmlValue)
  plain.text
}

paste("Ethereum was priced at", cryptoMC("ethereum"))

## [1] "Ethereum was priced at 208.90"

paste("Litecoin was priced at", cryptoMC("litecoin"))

## [1] "Litecoin was priced at 54.71"

Basic text functions in R

We will cover these using stringr as opposed to base R – stringr’s commands are much more consistent

• Every function in stringr can take a vector of strings for the first argument

Subsetting text

• Base R: Use substr() or substring()
• stringr: use str_sub()
  • First argument is a vector of strings
  • Second argument is the starting position (inclusive)
  • Third argument is that ending position (inclusive)

cat(str_wrap(str_sub(doc, 9896, 9929), 80))

## Citis net income was $13.5 billion

cat(str_wrap(str_sub(doc, 28900,29052), 80))

## Net income decreased 14%, mainly driven by lower revenues and lower loan loss
## reserve releases, partially offset by lower net credit losses and expenses.

Transforming text

• Commonly used functions:
  • tolower() or str_to_lower(): make the text lowercase
  • toupper() or str_to_lower(): MAKE THE TEXT UPPERCASE
  • str_to_title(): Make the Text Titlecase
• paste() to combine text
  • It puts spaces between by default
    • You can change this with the sep= option
  • If everything to combine is in 1 vector, use collapse= with the desired separator
  • paste0() is paste with sep=""

Examples: Case

sentence <- str_sub(doc, 9896, 9929)
str_to_lower(sentence)

## [1] "citis net income was $13.5 billion"

str_to_upper(sentence)

## [1] "CITIS NET INCOME WAS $13.5 BILLION"

str_to_title(sentence)

## [1] "Citis Net Income Was $13.5 Billion"

• The str_ prefixed functions support non-English languages as well

# You can run this in an R terminal! (It doesn't work in Rmarkdown though)
str_to_upper("Citis net income was $13.5 billion", locale='tr')  # Turkish

Examples: paste

# board is a list of director names
# titles is a list of the director's titles
paste(board, titles, sep=", ")

##  [1] "Michael L. Corbat, CEO"                                     
##  [2] "Michael E. O’Neill, Chairman"                               
##  [3] "Anthony M. Santomero, Former president, Fed (Philidelphia)" 
##  [4] "William S. Thompson, Jr., CEO, Retired, PIMCO"              
##  [5] "Duncan P. Hennes, Co-Founder/Partner, Atrevida Partners"    
##  [6] "Gary M. Reiner, Operating Partner, General Atlantic"        
##  [7] "Joan E. Spero, Senior Research Scholar, Columbia University"
##  [8] "James S. Turley, Former Chairman & CEO, E&Y"                
##  [9] "Franz B. Humer, Chairman, Roche"                            
## [10] "Judith Rodin, President, Rockefeller Foundation"            
## [11] "Robert L. Ryan, CFO, Retired, Medtronic"                    
## [12] "Diana L. Taylor, MD, Wolfensohn Fund Management"            
## [13] "Ernesto Zedillo Ponce de Leon, Professor, Yale University"  
## [14] "Robert L. Joss, Professor/Dean Emeritus, Stanford GSB"

cat(str_wrap(paste0("Citi's board consists of: ",
                    paste(board[1:length(board)-1], collapse=", "),
                    ", and ", board[length(board)], "."), 80))

## Citi's board consists of: Michael L. Corbat, Michael E. O’Neill, Anthony M.
## Santomero, William S. Thompson, Jr., Duncan P. Hennes, Gary M. Reiner, Joan E.
## Spero, James S. Turley, Franz B. Humer, Judith Rodin, Robert L. Ryan, Diana L.
## Taylor, Ernesto Zedillo Ponce de Leon, and Robert L. Joss.

Transforming text

• Replace text with str_replace_all()
  • First argument is text data
  • Second argument is what you want to remove
  • Third argument is the replacement
• If you only want to replace the first occurrence, use str_replace() instead

sentence

## [1] "Citis net income was $13.5 billion"

str_replace_all(sentence, "\\$13.5", "over $10")

## [1] "Citis net income was over $10 billion"

Transforming text

• Split text using str_split()
  • This function returns a list of vectors!
    • This is because it will turn every string passed to it into a vector, and R can’t have a vector of vectors
  • [[1]] can extract the first vector
• You can also limit the number of splits using n=
  • A bit more elegant solution is using str_split_fixed() with n=
    • Returns a character matrix (nicer than a list)

Example: Splitting text

paragraphs <- str_split(doc, '\n')[[1]]

# number of paragraphs
length(paragraphs)

## [1] 206

# Last paragraph
cat(str_wrap(paragraphs[206], 80))

## The total amount of securities authorized pursuant to any instrument defining
## rights of holders of long-term debt of the Company does not exceed 10% of the
## total assets of the Company and its consolidated subsidiaries. The Company
## will furnish copies of any such instrument to the SEC upon request. Copies of
## any of the exhibits referred to above will be furnished at a cost of $0.25 per
## page (although no charge will be made for the 2013 Annual Report on Form 10-
## K) to security holders who make written request to Citigroup Inc., Corporate
## Governance, 153 East 53 rd Street, 19 th Floor, New York, New York 10022. *
## Denotes a management contract or compensatory plan or arrangement. + Filed
## herewith.

Finding phrases in text

• How did I find the previous examples?

str_locate_all(tolower(doc), "net income")

## [[1]]
##        start    end
##  [1,]   8508   8517
##  [2,]   9902   9911
##  [3,]  16549  16558
##  [4,]  17562  17571
##  [5,]  28900  28909
##  [6,]  32197  32206
##  [7,]  35077  35086
##  [8,]  37252  37261
##  [9,]  40187  40196
## [10,]  43257  43266
## [11,]  45345  45354
## [12,]  47618  47627
## [13,]  51865  51874
## [14,]  51953  51962
## [15,]  52663  52672
## [16,]  52748  52757
## [17,]  54970  54979
## [18,]  58817  58826
## [19,]  96022  96031
## [20,]  96717  96726
## [21,]  99297  99306
## [22,] 188340 188349
## [23,] 189049 189058
## [24,] 201462 201471
## [25,] 456097 456106
## [26,] 460158 460167
## [27,] 460446 460455
## [28,] 460467 460476
## [29,] 475016 475025
## [30,] 475298 475307
## [31,] 545581 545590
## [32,] 554362 554371

Finding phrases in text

• 4 primary functions:
  1. str_detect(): Reports TRUE or FALSE for the presence of a string in the text
  2. str_count(): Reports the number of times a string is in the text
  3. str_locate(): Reports the first location of a string in the text
     • str_locate_all(): Reports every location as a list of matrices
  4. str_extract(): Reports the matched phrases
• All take a character vector as the first argument, and something to match for the second argument

Example: Finding phrases

• How many paragraphs mention net income in any case?

x <- str_detect(str_to_lower(paragraphs), "net income")
x[1:10]

##  [1] FALSE FALSE FALSE FALSE FALSE  TRUE FALSE FALSE  TRUE  TRUE

sum(x)

## [1] 13

• What is the most net income is mentioned in any paragraph

x <- str_count(str_to_lower(paragraphs), "net income")
x[1:10]

##  [1] 0 0 0 0 0 4 0 0 2 2

max(x)

## [1] 4

Example: Finding phrases

• Where is net income first mentioned in the document?

str_locate(str_to_lower(doc), "net income")

##      start  end
## [1,]  8508 8517

• First mention of net income
  • This function may look useless now, but it’ll be on of the most useful later

str_extract(str_to_lower(doc), "net income")

## [1] "net income"

R Practice

• Text data is already loaded, as if it was loaded using read_file()
• Try:
  • Subsetting the text data
  • Transforming the text data
    • To all upper case
    • Replacing a phrase
  • Finding specific text in the document
• Do exercises 1 through 3 in today’s practice file
  • R Practice
  • Shortlink: rmc.link/420r8

Finding patterns in the text (regex)

• Regular expressions, aka regex or regexp, are ways of finding patterns in text
• This means that instead of looking for a specific phrase, we can match a set of phrases
• Most of the functions we discussed accept regexes for matching
  • str_replace(), str_split(), str_detect(), str_count(), str_locate(), and str_extract(), plus their variants
• This is why str_extract() is so great!
  • We can extract anything from a document with it!

Regex example

• Breaking down an email
  1. A local name
  2. An @ sign
  3. A domain, which will have a . in it
• Local names can have many different characters in them
  • Match it with [:graph:]+
• The domain is pretty restrictive, generally just alphanumeric and .
  • There can be multiple . though
  • Match it with [:alnum:]+\\.[.[:alnum:]]+

# Extract all emails from the annual report
str_extract_all(doc,'[:graph:]+@[:alnum:]+\\.[.[:alnum:]]+')

## [[1]]
## [1] "shareholder@computershare.com" "shareholder@computershare.com"
## [3] "docserve@citi.com"             "shareholderrelations@citi.com"

Breaking down the example

• @ was itself – it isn’t a special character in strings in R
• \\. is just a period – we need to escape . because it is special in R
• Anything in brackets with colons, [: :], is a set of characters
  • [:graph:] means any letter, number, or punctuation
  • [:alnum:] means any letter or number
• + is used to indicate that we want 1 or more of the preceding element – as many as it can match
  • [:graph:]+ meant “Give us every letter, number, and punctuation you can, but make sure there is at least 1.”
• Brackets with no colons, [ ], ask for anything inside
  • [.[:alnum:]]+ meant “Give us every letter, number, and . you can, but make sure there is at least 1.”

Breaking down the example

• Let’s examine the output shareholder@computershare.com
• Our regex was [:graph:]+@[:alnum:]+\\.[.[:alnum:]]+
• Matching regex components to output:
  • [:graph:]+ $\Rightarrow$ shareholder
  • @ $\Rightarrow$ @
  • [:alnum:]+ $\Rightarrow$ computershare
  • \\. $\Rightarrow$ .
  • [.[:alnum:]]+ $\Rightarrow$ com

Useful regex components: Content

• There’s a nice cheat sheet here
  • More detailed documentation here
• Matching collections of characters
  • . matches everything
  • [:alpha:] matches all letters
  • [:lower:] matches all lowercase letters
  • [:upper:] matches all UPPERCASE letters
  • [:digit:] matches all numbers 0 through 9
  • [:alnum:] matches all letters and numbers
  • [:punct:] matches all punctuation
  • [:graph:] matches all letters, numbers, and punctuation
  • [:space:] or \s match ANY whitespace
    • \S is the exact opposite
  • [:blank:] matches whitespace except newlines

Example: Regex content

text <- c("abcde", 'ABCDE', '12345', '!?!?.', 'ABC123?', "With space", "New\nline")
html_df(data.frame(
  text=text,
  alpha=str_detect(text,'[:alpha:]'),
  lower=str_detect(text,'[:lower:]'),
  upper=str_detect(text,'[:upper:]'),
  digit=str_detect(text,'[:digit:]'),
  alnum=str_detect(text,'[:alnum:]')
))

Example: Regex content

text <- c("abcde", 'ABCDE', '12345', '!?!?.', 'ABC123?', "With space", "New\nline")
html_df(data.frame(
  text=text,
  punct=str_detect(text,'[:punct:]'),
  graph=str_detect(text,'[:graph:]'),
  space=str_detect(text,'[:space:]'),
  blank=str_detect(text,'[:blank:]'),
  period=str_detect(text,'.')
))

Useful regex components: Form

• [ ] can be used to create a class of characters to look for
  • [abc] matches anything that is a, b, c
• [^ ] can be used to create a class of everything else
  • [^abc] matches anything that isn’t a, b, or c
• Quantity, where x is some element
  • x? looks for 0 or 1 of x
  • x* looks for 0 or more of x
  • x+ looks for 1 or more of x
  • x{n} looks for n (a number) of x
  • x{n, } looks for at least n of x
  • x{n,m} looks for at least n and at most m of x
• Lazy operators
  • Append ? to any quantity operator to make it prefer the shortest match possible

Useful regex components: Form

• Position
  • ^ indicates the start of the string
  • $ indicates the end of the string
• Grouping
  • ( ) can be used to group components
  • | can be used within groups as a logical or
  • Groups can be referenced later using the position of the group within the regex
    • \\1 refers to the first group
    • \\2 refers to the second group
    • …

Example: Regex form (292 Real estate firms)

# Real estate firm names with 3 vowels in a row
str_subset(RE_names, '[AEIOU]{3}')

## [1] "STADLAUER MALZFABRIK"      "JOAO FORTES ENGENHARIA SA"

# Real estate firm names with no vowels
str_subset(RE_names, '^[^AEIOU]+$')

## [1] "FGP LTD"     "MBK PCL"     "MYP LTD"     "MCT BHD"     "R T C L LTD"

# Real estate firm names with at least 12 vowels
str_subset(RE_names, '([^AEIOU]*[AEIOU]){11,}')

## [1] "INTERNATIONAL ENTERTAINMENT"  "PREMIERE HORIZON ALLIANCE"   
## [3] "JOAO FORTES ENGENHARIA SA"    "OVERSEAS CHINESE TOWN (ASIA)"
## [5] "COOPERATIVE CONSTRUCTION CO"  "FRANCE TOURISME IMMOBILIER"  
## [7] "BONEI HATICHON CIVIL ENGINE"

# Real estate firm names with a repeated 4 letter pattern
str_subset(RE_names, '([:upper:]{4}).*\\1')

## [1] "INTERNATIONAL ENTERTAINMENT"  "CHONG HONG CONSTRUCTION CO"  
## [3] "ZHONGHONG HOLDING CO LTD"     "DEUTSCHE GEOTHERMISCHE IMMOB"

Why is regex so important?

• Regex can be used to match anything in text
  • Simple things like phone numbers
  • More complex things like addresses
• It can be used to parse through large markup documents
  • HTML, XML, LaTeX, etc.
• Very good for validating the format of text
  • For birthday in the format YYYYMMDD, you could validate with:
    • YYYY: [12][90][:digit:][:digit:]
    • MM: [01][:digit:]
    • DD: [0123][:digit:]

Cavaet: Regexes are generally slow. If you can code something to avoid them, that is often better. But often that may be infeasible.

Some extras

• While the str_*() functions use regex by default, they actually have four modes
  1. You can specify a regex normally
     • Or you can use regex() to construct more customized ones, such as regexes that operate by line in a string
  2. You can specify an exact string to match using fixed() – fast but fragile
  3. You can specify an exact string to match using coll() – slow but robust; recognizes characters that are equivalent
  4. You can ask for boundaries with boundary() such as words, using boundary("word")

Expanding usage

• Anything covered so far can be used for text in data
  • Ex.: Firm names or addresses in Compustat

# Compustat firm names example
df_RE_names <- df_RE %>%
  group_by(isin) %>%
  slice(1) %>%
  mutate(SG_in_name = str_detect(conm, "(SG|SINGAPORE)"),
         name_length = str_length(conm),
         SG_firm = ifelse(fic=="SGP",1,0)) %>%
  ungroup()

## Warning: package 'bindrcpp' was built under R version 3.5.1

df_RE_names %>%
  group_by(SG_firm) %>%
  mutate(pct_SG = mean(SG_in_name) * 100) %>%
  slice(1) %>%
  ungroup() %>%
  select(SG_firm, pct_SG)

## # A tibble: 2 x 2
##   SG_firm pct_SG
##     <dbl>  <dbl>
## 1       0  0.369
## 2       1  4.76

Expanding usage

library(DT)
df_RE_names %>%
  group_by(fic) %>%
  mutate(avg_name_length = mean(name_length)) %>%
  slice(1) %>%
  ungroup() %>%
  select(fic, avg_name_length) %>%
  arrange(desc(avg_name_length), fic) %>%
  datatable(options = list(pageLength = 5))

R Practice 2

• This practice explores the previously used practice data using regular expressions for various purposes
• Do exercises 4 and 5 in today’s practice file
  • R Practice
  • Shortlink: rmc.link/420r8

Readability

• Thanks to the quanteda package, readability is very easy to calculate in R
  • Use the textstat_readability() function
• There are many readability measures, however
  • Flesch Kinkaid: A measure of readability developed for the U.S. Navy to ensure manuals were written at a level any 15 year old should be able to understand
  • Fog: An index that was commonly used in business and publishing
  • Coleman-Liau: An index with a unique calculation method

Readability: Flesch Kincaid

$\displaystyle { 206.835 - 1.015 \left(\frac{\#~words}{\#~sentences}\right) - 84.6\left(\frac{\#~syllables}{\#~words}\right) }$

• A score generally below 100
  • Higher is more readable
  • Conversational English should be around 80-90
  • A JC or poly graduate should be able to read anything 50 or higher
  • A Bachelor’s degree could be necessary for anything below 30

library(quanteda)

## Warning: package 'quanteda' was built under R version 3.5.1

textstat_readability(doc, "Flesch.Kincaid")

##   document Flesch.Kincaid
## 1    text1       17.56528

Readability: Fog

$\displaystyle { \begin{aligned} & \left[ Mean(Words~per~sentence) +\right.\\ &\left.(\%~of~words~>3~syllables)\right] \times 0.4 \end{aligned} }$

• An approximate grade level required for reading a document
  • A JC or poly graduate should read at a level of 12
    • New York Times articles are usually around 13
  • A Bachelor’s degree holder should read at 17

textstat_readability(doc, "FOG")

##   document      FOG
## 1    text1 21.63388

Readability: Coleman-Liau

$\displaystyle { 5.88\left(\frac{\#~letters}{\#~words}\right)-29.6\left(\frac{\#~sentences}{\#~words}\right)-15.8 }$

• Provides an approximate grade level like Fog, on the same scale as Fog

textstat_readability(doc, "Coleman.Liau")

##   document Coleman.Liau
## 1    text1     29.03967

Converting text to words

• Tidy text is when you have when token per document per row, in a data frame
• Token is the unit of text you are interested in
  • Words: “New”
  • Phrases: “New York Times”
  • Sentences: “The New York Times is a publication.”
  • etc.
• The tidytext package can handle this conversion for us!
  • Use the unnest_tokens() function
  • Note: it also converts to lowercase. Use the option to_lower=FALSE to avoid this if needed

# Example of "tokenizing"
library(tidytext)
df_doc <- data.frame(ID=c("0001104659-14-015152"), text=c(doc),
                     stringsAsFactors = F) %>%
  unnest_tokens(word, text)
# word is the name for the new column
# text is the name of the string column in the input data

The details

• tidytext uses the tokenizers package in the backend to do the conversion
  • You can call that package directly instead if you want to
• Available tokenizers include: (specify with token=)
  • “word”: The default, individual words
  • “ngram”: Collections of words (default of 2, specify with n=)
  • A few other less commonly used tokenizers

Word case

• Why convert to lowercase?
• How much of a difference is there between “The” and “the”?
  • “Singapore” and “singapore” – still not much difference
  • Only words like “new” versus “New” matter
    • “New York” versus “new yorkshire terrier”
• Benefit: We get rid of a bunch of distinct words!
  • Helps with the curse of dimensionality

The Curse of dimensionality

• There are a lot of words
• A LOT OF WORDS
• At least 171,476 according to Oxford Dictionary
• What happens if we make a matrix of words per document?

Stopwords

• Stopwords – words we remove because they have little content
  • the, a, an, and, …
• Also helps with our curse a bit – removes the words entirely
• We’ll use the tm package to remove stopwords
  • Uses a mix of SMART and Snowball stemmer under the hood

# get a list of stopwords
library(stopwords)
stop_en <- stopwords("english")  # Snowball English
paste0(length(stop_en), " words: ", paste(stop_en[1:5], collapse=", "))

## [1] "175 words: i, me, my, myself, we"

stop_SMART <- stopwords(source="smart")  # SMART English
paste0(length(stop_SMART), " words: ", paste(stop_SMART[1:5], collapse=", "))

## [1] "571 words: a, a's, able, about, above"

stop_fr <- stopwords("french")  # Snowball French
paste0(length(stop_fr), " words: ", paste(stop_fr[1:5], collapse=", "))

## [1] "164 words: au, aux, avec, ce, ces"

Applying stopwords to a corpus

• When we have a tidy set of text, we can just use dplyr for this!
  • dplyr’s anti_join() function is like a merge, but where all matches are deleted

df_doc_stop <- df_doc %>%
  anti_join(data.frame(word=stop_SMART, stringsAsFactors = F))

## Joining, by = "word"

nrow(df_doc)

## [1] 128728

nrow(df_doc_stop)

## [1] 74985

Converting to term frequency

terms <- df_doc_stop %>%
  count(ID, word, sort=TRUE) %>%
  ungroup()
total_terms <- terms %>% 
  group_by(ID) %>% 
  summarize(total = sum(n))
tf <- left_join(terms, total_terms) %>% mutate(tf=n/total)

## Joining, by = "ID"

tf

## # A tibble: 5,543 x 5
##    ID                   word          n total      tf
##    <chr>                <chr>     <int> <int>   <dbl>
##  1 0001104659-14-015152 citi        826 74985 0.0110 
##  2 0001104659-14-015152 2013        743 74985 0.00991
##  3 0001104659-14-015152 credit      704 74985 0.00939
##  4 0001104659-14-015152 citis       660 74985 0.00880
##  5 0001104659-14-015152 risk        624 74985 0.00832
##  6 0001104659-14-015152 december    523 74985 0.00697
##  7 0001104659-14-015152 financial   513 74985 0.00684
##  8 0001104659-14-015152 31          505 74985 0.00673
##  9 0001104659-14-015152 loans       495 74985 0.00660
## 10 0001104659-14-015152 assets      488 74985 0.00651
## # ... with 5,533 more rows

Sentiment

• Sentiment works similarly to stopwords, except we are identifying words with specific, useful meanings
  • We can grab off-the-shelf sentiment measures using get_sentiments() from tidytext

get_sentiments("afinn") %>%
  group_by(score) %>%
  slice(1) %>%
  ungroup()

## # A tibble: 11 x 2
##    word         score
##    <chr>        <int>
##  1 bastard         -5
##  2 ass             -4
##  3 abhor           -3
##  4 abandon         -2
##  5 absentee        -1
##  6 some kind        0
##  7 aboard           1
##  8 abilities        2
##  9 admire           3
## 10 amazing          4
## 11 breathtaking     5

get_sentiments("bing") %>%
  group_by(sentiment) %>%
  slice(1) %>%
  ungroup()

## # A tibble: 2 x 2
##   word    sentiment
##   <chr>   <chr>    
## 1 2-faced negative 
## 2 a+      positive

Sentiment

get_sentiments("nrc") %>%
  group_by(sentiment) %>%
  slice(1) %>%
  ungroup()

## # A tibble: 10 x 2
##    word        sentiment   
##    <chr>       <chr>       
##  1 abandoned   anger       
##  2 abundance   anticipation
##  3 aberration  disgust     
##  4 abandon     fear        
##  5 absolution  joy         
##  6 abandon     negative    
##  7 abba        positive    
##  8 abandon     sadness     
##  9 abandonment surprise    
## 10 abacus      trust

get_sentiments("loughran") %>%
  group_by(sentiment) %>%
  slice(1) %>%
  ungroup()

## # A tibble: 6 x 2
##   word           sentiment   
##   <chr>          <chr>       
## 1 abide          constraining
## 2 abovementioned litigious   
## 3 abandon        negative    
## 4 able           positive    
## 5 aegis          superfluous 
## 6 abeyance       uncertainty

Merging in sentiment data

tf_sent <- tf %>% left_join(get_sentiments("loughran"))

## Joining, by = "word"

tf_sent[1:5,]

## # A tibble: 5 x 6
##   ID                   word       n total      tf sentiment  
##   <chr>                <chr>  <int> <int>   <dbl> <chr>      
## 1 0001104659-14-015152 citi     826 74985 0.0110  <NA>       
## 2 0001104659-14-015152 2013     743 74985 0.00991 <NA>       
## 3 0001104659-14-015152 credit   704 74985 0.00939 <NA>       
## 4 0001104659-14-015152 citis    660 74985 0.00880 <NA>       
## 5 0001104659-14-015152 risk     624 74985 0.00832 uncertainty

tf_sent[!is.na(tf_sent$sentiment),][1:5,]

## # A tibble: 5 x 6
##   ID                   word              n total      tf sentiment  
##   <chr>                <chr>         <int> <int>   <dbl> <chr>      
## 1 0001104659-14-015152 risk            624 74985 0.00832 uncertainty
## 2 0001104659-14-015152 loss            267 74985 0.00356 negative   
## 3 0001104659-14-015152 losses          265 74985 0.00353 negative   
## 4 0001104659-14-015152 approximately   232 74985 0.00309 uncertainty
## 5 0001104659-14-015152 regulatory      216 74985 0.00288 litigious

Summarizing document sentiment

tf_sent %>%
  spread(sentiment, tf, fill=0) %>%
  select(constraining, litigious, negative, positive, superfluous, uncertainty) %>%
  colSums()

## constraining    litigious     negative     positive  superfluous 
##  0.013242649  0.020750817  0.034780289  0.007054744  0.000373408 
##  uncertainty 
##  0.025325065

visualizing sentiment

Visualizing a document as a word cloud

• quanteda also provides an easy way to make a word cloud
  • textplot_wordcloud()
• There are also the wordcloud and wordcloud2 packages for this

corp <- corpus(df_doc_stop, docid_field="ID", text_field="word")
textplot_wordcloud(dfm(corp), color = RColorBrewer::brewer.pal(10, "RdBu"))

Another reason to use stopwords

• Without removing stopwords, the word cloud shows almost nothing useful

corp <- corpus(df_doc, docid_field="ID", text_field="word")
textplot_wordcloud(dfm(corp), color = RColorBrewer::brewer.pal(10, "RdBu"))

R Practice 3

• Using the same data as before, we will explore
  • Readability
  • Sentiment
  • Word clouds
• Note: Due to missing packages, you will need to run the code in RStudio, not in the DataCamp light console
• Do exercises 6 through 8 in today’s practice file
  • R Practice
  • Shortlink: rmc.link/420r8

For next week

• For next week:
  • Finish the third assignment
    • Submit on eLearn
  • Datacamp
    • Do the assigned chapter on text analysis
  • Keep working on the group project

Packages used for these slides

• kableExtra
• knitr
• magrittr
• quanteda
• RColorBrewer
• RCurl
• readtext
• revealjs
• tidytext
• tidyverse
  • dplyr, readr, stringr
• XML

Custom code

library(knitr)
library(kableExtra)
html_df <- function(text, cols=NULL, col1=FALSE, full=F) {
  if(!length(cols)) {
    cols=colnames(text)
  }
  if(!col1) {
    kable(text,"html", col.names = cols, align = c("l",rep('c',length(cols)-1))) %>%
      kable_styling(bootstrap_options = c("striped","hover"), full_width=full)
  } else {
    kable(text,"html", col.names = cols, align = c("l",rep('c',length(cols)-1))) %>%
      kable_styling(bootstrap_options = c("striped","hover"), full_width=full) %>%
      column_spec(1,bold=T)
  }
}

cryptoMC <- function(name) {
  if (exists(name)) {
    get(name)
  } else{
  html <- getURL(paste('https://coinmarketcap.com/currencies/',name,'/',sep=''))
  xpath <- '//*[@id="quote_price"]/span[1]/text()'
  doc = htmlParse(html, asText=TRUE)
  plain.text <- xpathSApply(doc, xpath, xmlValue)
  assign(name, gsub("\n","",gsub(" ", "", paste(plain.text, collapse = ""), fixed = TRUE), fixed = TRUE),envir = .GlobalEnv)
  get(name)
  }
}

# Loads line-by-line by default
# This makes it document-by-document
library(textreadr)
df2 <- read_dir("G:/2014/2014/") %>%
  group_by(document) %>%
  mutate(text=paste(content, collapse="\n")) %>%
  select(document,text)
  slice(1) %>%
  ungroup()

Custom code

# Create a plot of the top words by sentiment
tf_sent %>%
  filter(!is.na(sentiment)) %>%
  group_by(sentiment) %>%
  arrange(desc(n)) %>%
  mutate(row = row_number()) %>%
  filter(row < 10) %>%
  ungroup() %>%
  mutate(word = reorder(word, n)) %>%
  ggplot(aes(y=n, x=word)) + geom_col() + theme(axis.text.x = element_text(angle=90, hjust=1)) + 
  facet_wrap(~sentiment, ncol=3, scales="free_x")