Frequency Analysis

• frequency ~ measure of relevance
• bag of words approach
• simple
• powerful

Key Figures of Texts

wc *.txt    # count number of lines, words, characters

Word Occurrences

show in context

egrep -ir "computational" folder/       # search in all files in folder, ignore case

# common egrep options:
# -i            search case-insensitive
# -r            search recursively in all subfolders
# --colour      highlight matches
# --context 2   show 2 lines above/below match

count words

egrep -ic "big data" *.txt      # count across all txt-files, ignore case

Word Frequencies

steps of the algorithm

1. split text into one word per line (tokenize)
2. sort words alphabetically
3. count how often each word appears

# piping steps to get word frequencies
cat text.txt | tr " " "\n" | sort | uniq -c | sort -h > wordfreq.txt

# explanation of individual steps:
tr " " "\n"     # replace spaces with newline 
sort -h         # sort lines alphanumerically
uniq -c         # count repeated lines

Word Frequencies

• absolute frequency
• relative frequency
  • = n_occurrences / n_total_words
  • independent of size
• statistical validation of variation
  • significance tests between corpora

Convert Stats into Dataset

• convert to .tsv file
• useful for further processing
  • e.g., import in Excel

# convert word frequencies into tsv-file
# additional step: replace a sequence of spaces with a tabulator
cat text.txt | tr " " "\n" | sort | uniq -c  | sort -h | \
tr -s " " "\t"  > test.tsv  

In-class: Matching and counting

1. Print the following sentence in your command line using echo.

   echo "There are a few related fields: NLP, computational linguistics, and computational text analysis."

2. How many words are in this sentence? Use the pipe operator to combine the command above with wc.

3. Match the words computational and colorize its occurences in the sentence using egrep.

4. Get the frequencies of each word in this sentence using tr and other commands.

Common Preprocessing

Refining results with

• lowercasing
• replace symbols
• join lines
• trimming header + footer
• splitting into multiple files
• using patterns to remove/extract parts 📅

Lowercasing

reduce word forms

echo "ÜBER" | tr "A-ZÄÖÜ" "a-zäöü"  # fold text to lowercase

Removing and Replacing Symbols

echo "3x3" | tr -d "[:digit:]"      # remove all digits 
cat text.txt | tr -d "[:punct:]"    # remove punctuation like .,:;?!- 

tr "Y" "Z"                          # replace any Y with Z

Standard Preprocessing

save a preprocessed document

# lowercase, no punctuation, no digits
cat speech.txt | tr "A-ZÄÖÜ" "a-zäöü" | \
tr -d "[:punct:]" | tr -d "[:digit:]" > speech_clean.txt

Join Lines

cat test.txt | tr -s "\n" " "   # replace newlines with spaces

Trim Lines

cat -n text.txt         # show line numbers
sed "1,10d" text.txt    # remove lines 1 to 10

Splitting Files

# splits file at every delimiter into a stand-alone file
csplit huge_text.txt  "/delimiter/" {*}

Check Differences between Files

sanity check after modification

# show differences side-by-side and only differing lines
diff -y --suppress-common-lines text_raw.txt text_proc.txt

Where there is a shell,
there is a way. 👍

Organizing Code

• Git tracks file changes and allows for version management
• GitHub is a popular hosting platform based on Git
  • share code and collaborate
  • repository = project folder

🤓 Published code and data are parts of the endeavour of open science.

In-class: Getting ready

1. Change into your local copy of the GitHub course repository KED2022 and update it with git pull. When you haven’t cloned the repository, follow section 5 of the installation guide .

   You find some party programmes (Grüne, SP, SVP) in materials/data/swiss_party_programmes/txt. The programmes are provided in plain text which I have extracted from the publicly available PDFs.

2. Have a look at the content of some of these text files using more.

In-class: Analyzing Swiss Party Programmes I

1. Compare the absolute frequencies of single terms or multi-word expressions of your choice (e.g., Ökologie, Sicherheit, Schweiz)…

   • across parties
   • historically within a party

   Use the file names as filter to get various aggregation of the word counts.

2. Pick terms of your interest and look at their contextual use by extracting relevant passages. Does the usage differ across parties or time?

Share your insights with the class using Etherpad.

In-class: Analyzing Swiss Party Programmes II

1. Convert the word frequencies per party into a tsv dataset. Compute the relative word frequency instead of the absolute frequency using any spreadsheet software (e.g. Excel). Are your conclusions still valid after accounting for the size?
2. Can you refine the results with further preprocessing of the data?
3. What is the size of the vocabulary of this data collection (number of unique words)?

Pro Tip 🤓: Use egrep to look up commands in the .md course slides

Additional Resources

When you look for useful primers on Bash, consider the following resources:

• Tutorial Basic Text Analysis by W. Turkel
• Tutorial Pattern Matching + KWIC by W. Turkel