Repository for work related to Mike Taylor and Rebecca Ballard's "Native Newspaper" project.
- Use
atom-harvester.pyto harvest atom data from the Chronicling America website for the dates we are interested in (1 Nov 1890 - 31 March 1891). - Use
atom-reader.pyto read each atom file and create a data dictionary. - Use
ocr-harvester.pyto download the.HTMLof the OCR data for each newspaper page for the dates we are interested in. - Use
ocr-checker.pyto verify that the number of downloaded HTML pages matches what was expected. - Use
ocr-extractor.pyto extract the OCR from the HTML pages. - Use
ocr-combiner.pyto combine the individual pages of a newspaper issue into a singletxtfile.
N.B. At first, atom-reader.py got the link to the HTML page for the OCR. When I got to extracting the OCR with ocr-extractor.py, I realized that I could have downloaded the OCR directly without going through the HTML. For that reason, I rewrote atom-reader.py to produce a different link to the OCR. If I were to go through the whole process again, ocr-extractor.py would be unnecessary. That said, the data we have were actually created by extracting the OCR from the HTML. Sticking with this process saved us from having to re-harvest all 76k pages.
- Use
search.pyto find frequencies of key terms. - Topic model the contents of the
combined-ocrfolder using MALLET. (Unfortunately, I don't seem to have kept a record of the command. I did 100 topics, stop words removed, 1000 iterations [default] and optimization set at, I believe, 100. I am also 95% certain that I set a seed of1so I could replicate the results.) - Use
topic-analyzer.pyto isolate topic 70 from the topic model. - Use
visualize-terms.pyto create graphs of weekly appearances of 'ghost dance', 'pine ridge', or 'wounded knee' - Use
find_2jan.pyto find all issues published on 1891-01-02 and copy them to a separate folder.
This script downloads 3838 atom files from the Chronicling America website based on a search query that is formulated for the dates we are interested in. The atom files each contain 20 results, each of which is a single page of a newspaper within the timeframe. The atom files are saved as xml in the atom-data folder.
This script processes the atom files in the atom-data folder. It reads the entries in the atom file and outputs the results to data-dictionary.tsv.
This script finds every newspaper issue in the combined-ocr folder that has a publication date of 1891-01-2 and copies that issue to the 2-jan-issues folder.
This script takes the data from newspapers.txt and converts it into a tsv with columns for city and state for the newspapers. The output is newspapers_locations.tsv.
This script creates two sets: one is the OCR HTML files that were expected to be downloaded from the information in data-dictionary.tsv; the other is the OCR HTML files that were actually downloaded using ocr-harvester.py.
I created this script because data-dictionary.tsv was 76,748 lines long, but I only got 74,978 HTML files downloaded from LOC. After a number of ways of checking (which resulted in this script), I verified that there were duplicate lines in data-dictionary.tsv and that the number of duplicates meant that the number of files I was actually expecting to download and the number that I did download matched.
This script crawls the data in the ocr-txt folder and combines the individual pages of a newspaper issue into a single txt file. It outputs its results to the combined-ocr folder.
This script extracts the text of the OCR from the HTML that was downloaded with ocr-harvester.py. It saves its output to the ocr-txt folder.
This script uses the information in data-dictionary.tsv to download .txt OCR files from the Chronicling America website and save them in the ocr-txt folder. It also outputs a set of all newspapers in the data set, saving it to newspapers.txt.
This script uses regex to search for key terms within the corpus contained in the combined-ocr folder. It saves the output (including number of hits and strings that hit) to search-results.tsv. It also saves the total counts for each term to search-counters.txt.
This script takes the results of the 100-topic topic model and isolates topic 70, which is the "Wounded Knee" topic. It sets the topic-composition of the document (a single newspaper issue) to 0 if the composition is below 0.01 (1%). It also rounds the topic composition to 4 decimal places. It saves this output to tm_wk_results.tsv. As a secondary action, it saves newspaper issues that are composed of 0.01 or higher in topic 70 to the tm-subset folder.
This script finds all the locations of the tar files for the Chronicling America website. It would be a first step if we wanted to bulk download all the data in the data set. It saves its output to tarfiles.txt. This was the first step I took in this project and I subsequently abandoned this line of work. An intermediary step is the creation of chronicling-atom.txt.
This script takes the output of search.py and creates graphs for the weekly appearance of key terms in the corpus. It saves its output to the images folder. This script also creates a .tsv of "ghost dance" mentions that we provided to PMLA for the publication of the article.
This folder has the output of the find_2jan.py script. It is a collection of all newspaper issues that were published on 1891-01-02.
This folder has the output of the atom-harvester.py script. It is a collection of 3,838 atom/XML files.
This folder has the output of the ocr-combiner.py script. It is a collection of individual issues of newspaper issues.
This folder has the output of the visualize-terms.py script. It is a collection of maps of key terms over time.
This folder has the output of the ocr-harvester.py script. It is a collection of 76k HTML files downloaded from the Chronicling America website. N.B. If re-running the scripts, this folder would be unnecessary.
This folder has the output of the ocr-extractor.py script. It is a collection of 76k text files that were extracted from the data in the ocr-html folder.
This folder has the output of the topic-analyzer.py script. It is a collection of newspaper issues that are composed of 0.01 or higher of topic 70, the "Wounded Knee" topic.
This is an output of url-grabber.py and is an intermediary step to getting all the tar files from the site.
This tsv is the output of atom-reader.py. For each entry in the atom data, it lists
- newspaper
- date published (in ISO format)
- the page number of the image, formatted as
seq-4where4is the page number - the direct link to Chronicling America for that page
- the link to the OCR for that page This file is used for step 3 in the order of operations.
This tsv is one of the outputs of visualize-terms.py. It has the data that is used to produce the graph of "ghost dance" mentions so graphs can be made outside of Python.
This is an output of newspaper_tsv_prep.py and is a list of all the newspapers in our targeted corpus as well as their city and states.
This is an output of ocr-harvester.py, and is a list of all the newspapers in our targeted corpus.
This is the output of ocr-checker.py. It shows which items appear more than once in data-dictionary.tsv.
This markdown file contains preliminary notes that I created while trying to understand how to formulate a proper search query to use in harvesting-chronicling.py. In it, I also sketch out the different scripts that will be needed to accomplish the work that I need to do.
This is an output of search.py. It is a list of total number of hits for the regex terms in search.py.
This is an output of search.py. It is a list of all newspaper issues and the number of hits for each regex term in search.py and the strings that triggered those hits.
This is an output of url-grabber.py and is a list of direct downloads for the tar files for each part of the Chronicling America data set. This could be used if we wanted to bulk download all of the data from the site.
This is an output of topic-analyzer.py and is a tsv of the results of the topic model. It lists the newspaper, the date of the issue, and the percentage of that document composed of topic 70, which is the "Wounded Knee" topic. The percentage was rounded to 4 decimal points, and anything below 0.01 was set to 0.