Intro

The goal of these two Python scripts (text_processing_simple.py and text_processing_filtered.py) is to preprocess samples of text data in order to parse out ‘significant language.’ How ‘significant langauge’ is defined depends on which script you use.

text_processing_simple.py:

Basic Idea

This script performs a simple analysis that will return the top 200 most frequently occurring phrases (from one to three words long) from your sample, and how many documents (samples) they occurred on.

Details

Inputs

• A .txt file containing your sample of texts, each text in its own row, and no header. The texts should be cleaned so that each word is delimited by a specific character (EX: ‘|’). The file can have an ID column in the first column (so then texts are in the second column), or just a single column containing the texts
• Parameters for setting threshold on which single words to keep:
  • Minimum threshold = min number of texts a word has to appear in to be kept
  • Maximum threshold = max number of texts a word is allowed to appear in to be kept

Outputs

• ‘top200_trigrams.txt’ is outputted to your working directory; this contains the top 200 most commonly occurring tokens and their document frequency
• Intermediate files are outputted to wherever the original .txt file was stored. These are .txt files that contain:
  • Doc frequencies for all single words in the original corpus
  • Doc frequencies for all uni, bi, and trigrams in the core language body of the corpus
  • The trigram versions of the original texts
• A log file called ‘simple.log’ is outputted to your working directory

text_processing_filtered.py:

Basic Idea

This script performs a more complex analysis by comparing your target sample with a non-target sample in order to filter out ‘meaningless’ language (aka language you are not interested in because they occur with the same incidence on your target and non-target samples). The script returns a list of all phrases that meet given thresholds to be considered ‘significant’, ordered by significance. These thresholds are:

• minimum docnum = minimum number of documents in your target sample a phrase has occurred on; given by user input
• multiplier = how many more times a phrase has occurred on documents in your target sample vs documents in your non-target sample; default is 2
• p-value = maximum p-value from calculating a t-test on the occurrence of a phrase on your target sample vs your non-target sample; default is 0.25

1. Texts from both samples are reduced to bodies of single words; rare words are kicked out.
2. Then, t-tests are conducted between the two samples on the single words retained from (1)
3. A generous p-value threshold is used to omit words that occur with similar frequency on both samples; these words are deleted from both samples. (Here we’re just trying to kick out words that are obviously irrelevant, not trying to draw any inferences)
4. Trigrams are created on both reduced samples; this performs faster than just creating trigrams on the full raw text since there are ‘holes’ everywhere that can be skipped over
5. t-tests are conducted between the two trigram’d samples created from (4) and results are outputted.

Details

Inputs

• A .txt file containing your sample of your target texts, each text in its own row, and no header. The texts should be cleaned so that each word is delimited by a specific character (EX: ‘|’). The file can have an ID column in the first column (so then texts are in the second column), or just a single column containing the texts
• A .txt file containing your filter sample of texts, same format as above.
• Parameters for setting threshold on which single words to keep:
  • Minimum threshold = min number of texts a word has to appear in to be kept

Outputs

• ‘top_trigrams.txt’ is outputted to your working directory; this contains:
  • col 1: tokens from the core language body of the corpus
  • col 2: p-values from conducting t-tests comparing token counts on the target sample vs the filter sample
  • col 3: document frequencies for the target texts
  • col 4: document frequencies for the filter texts
• Intermediate files are outputted to wherever the original .txt file was stored. These are .txt files that contain:
  • Doc frequencies for all single words in the original corpus for each sample
  • t-test results for comparing single words between the two samples
  • Doc frequencies for all uni, bi, and trigrams in the core language body of the corpus for each sample
  • The trigram versions of the original texts for each sample

Notes

• Three-word phrases are reduced so that the middle word is a free word and is represented by a dash (EX: ‘cats are animals’ becomes ‘cats - animals’)
• Default file encoding is cp1252 in the scripts due to usage with files from Windows applications. Default encoding is utf-8 in the modules

How to Use

Both scripts can be run from the command line:

python text_processing_filtered.py

They will prompt for user inputs when needed.

There is an example file containing public Congress transcripts in sample_texts that you can run the simple script on. Haven’t put up an appropriate pair of sample texts for the filtered script yet.

To do’s

• Integrate scripts somehow - seems redundant to have two separate scripts, when functionality and inputs are similar
• Get the file encoding to be a user input also
• Introduce steps prior to these scripts to help clean/standardize text into the desired format
• Add sample texts/examples to demonstrate scripts on
• Look into possible ways to boost performance
• Look at adding other parameters as possible user inputs