This repository accompanies the semester project by M. Sideri and J. Jordan for the project module course 'Mining Opinions and Argument' held at the University of Potsdam in the winter semester 2023/24.
Edited texts are located in the editing directory, separated into subdirectories by annotator. The individual
subdirectories also contain plain text files listing text IDs of texts used for the survey.
Prepared survey data, and raw and processed survey replies are located in the survey directory.
This project requires the Potsdam Argumentative Microtext Corpus.
The required corpus files are available at https://github.com/peldszus/arg-microtexts.
Download the corpus directory found in the linked repository and place it in the root directory of this project.
This project's code is written with the assumption that the corpus data is accessible at the path corpus/.
Python library requirements are listed in requirements.txt.
As the first step, run corpus_extractor.py. This script extracts various information from the corpus and creates a
JSON file, extracted_db.json, allowing easy access to this information for further use. Among others, it contains
abstracted linearization strategies (as described in Peldszus & Stede 2015),
central claim positions and argGraph node depths. See comments in corpus_extractor.py for specifics.
All further scripts use extracted_db.json.
corpus_stats.py calculates and prints various frequency statistics, mainly reproducing those presented in Peldszus & Stede 2015.
by_stance.py calculates linearization strategy statistics, separated by text main stance, reproducing Peldszus & Stede 2015
in more detail. The script can be run to show the calculated statistics.
unit_checking.py contains functions to identify split, sub-clause text units and anaphora, as well as some utility
functions. While attempts to use these for automation were abandoned, this script provides further insight into the
corpus. The script can be run to show the various features examined.
unit_bigram_ex.py contains an experimental attempt at identifying inseparable unit bigrams, which was soon abandoned
as it showed that the corpus data is too inconsistent to automate this. The script can be run to show the exploration
for a single text (b001).
Various scripts with 'explore'/'exploration' in their names were created as part of the corpus exploration, and while
they were not used further, they are left here in their original form for work documentation.
shuffle_test.py contains an attempt to automatically produce all linear order variants of a corpus text. This was
abandoned in favor of manual inspection and preliminary exclusion of non-viable order variants. The script can be run to
show the indiscriminately shuffled variants of a single text (b009).
editing_util.py contains an utility function to quickly create plain text files with units separated into lines and
annotations of unit indices and unit stances. These plain text files are to be used for manual rearrangement and editing.
The script can be run to show the resulting format and reproduce the first step of the editing procedure for the survey.
survey_util.py contains all functions used to prepare the survey and to process the raw survey replies.
survey_util.py can be run to reproduce the preparation and result processing for the survey.
Much of this script relies on raw text files that list corpus text IDs, with IDs separated by newline in the files. The
read_texts_list() function reads these files and returns their content as list.
The functions copy_original_texts() and copy_edited_texts() are used to conveniently copy original text files from
the corpus and edited text files from the editing directory to the survey directories for further use.
survey_texts_info() is used check the balance of originally claim-first and -last texts to be used for the survey,
based on manually created text list files. It also provides information about the topics and main stances of the listed
texts. survey_texts_frequencies() expresses the information from the former function as a dict of counts.
get_original_first() and get_original_last() are utility functions used to conveniently get lists of text IDs
based on their original order being claim-first or -last.
As there were more originally claim-first texts in our initial selection of texts to edit, create_balanced_text_list()
was used to assure that a selected subset of these texts stored as a text ID list file correctly matches the number of
originally claim-last files to be edited. It also saves a text ID list file to disk when the selection is balanced,
which contains the IDs of all texts to be used for the survey.
build_survey_pairs() creates a dict containing both the original texts and their edited versions for further
processing. survey_processing_pairs() creates a similar dict, which additionally contains the number of each text
on the survey form and table data retrieved after the survey period.
pairs_to_csv() creates a CSV file with the pairs of edited and unedited texts to be used in the survey.
pairs_to_simple_list() creates a plain text file with the pairs of edited and unedited texts, ordered as to be
presented on the survey form. The format in the saved file allows for convenient copying of the text pairs into the
Google Forms editing interface for quick addition of choices.
process_survey_answers() processes the retrieved raw survey data from CSV to a boolean format for further processing
and analysis. Result is returned as Pandas DataFrame.
survey_cross_counts() uses the processed survey data to count participant choices, separated by edited/unedited and
claim-first/-last. Result is returned as Pandas DataFrame.
Multiple files created during survey preparation and processing are left in the survey directory for work
documentation purposes.