This repository contains the code related to the paper "Mining Italian Short Argumentative Texts" (I. Namor, P.Totis, S. Garda, M.Stede). It is a fork of the project Evidence graphs for parsing argumentation structure by A.Peldszus.

We extend the original work to run on an Italian corpus of argumentative texts.

This code runs in Python 3.6. It is recommended to install it in a separate virtual environment. Here are installation instructions for an Ubuntu 18.04 linux:

# basics
sudo apt install virtualenv python3.6-dev
# for lxml
sudo apt install libxml2-dev libxslt1-dev
# for matplotlib
sudo apt install libpng-dev libfreetype6-dev
# for graph plotting
sudo apt install graphviz

Install all required python libaries in the environment and download the language models required by the spacy library.

make install-requirements
make download-spacy-data-de
make download-spacy-data-en
make download-spacy-data-it

After that you need to download all the models needed for matetools and place them in mateparser/models:

1. parser,tagger and morphology (here)
2. lemmatizer (here)

Make sure all the tests pass.

make test

Run a (shortened and simplified) minimal experiment, to see that everything is working:

make run-minimal-en

You should (see last lines of the output) get an average macro F1 of the base classifiers similar to: (cc ~= 0.846, ro ~= 0.758, fu ~= 0.745, at ~= 0.705).

Evaluate the results, which have been written to data/:

env/bin/python src/experiments/eval_minimal.py --lang en

You should (see first lines of the output) get an average macro F1 for the decoded results similar to: (cc ~= 0.861, ro ~= 0.771, fu ~= 0.754, at ~= 0.712).

Run one of the following:

make run-complete-en
make run-complete-de
make run-complete-it

It is possible to add the (SE annotations) to the (micro-texts) corpus.

In order to to so it is needed to first parse the annotations to match the actual ADUs found in the micro-texts corpus.

env/bin/python src/my_scripts/parse_set.py ./resources/adu2setypes.pkl

This will created a compressed dictionary containing all micro-texts ADUs with their corresponding SE types. In order to avoid mismatching with the ADUs representation in the features_text.py module, all the ADUs are in one line with no space.

Then it is possible to use

env/bin/python src/my_scripts/run_minimal_with_se_feat.py --lang de

to asses the impact of such features.

In this setting the micro-text corpus is the one provided by Andreas found in the folder argmin/data/corpus/german/arg. Therefore, you do not need to pass any specific path to the corpus.

env/bin/python src/my_scripts/files_to_pred_and_viz.py --out <folder> --model <path to pre-trained model>

where --out specify the path to a folder (if it does not exist it will creat it) and for --model I suggest to use model the last model produced with run_minimal.py, so it should be data/models/m112de-test-adu-full-noop__<hash>__4.

If you have in mind some interasting cases that you wish to check directly, instead of going through the all corpus you can pass them with --text-ids (comma separated list). So, for instance, if you want to see what happens only for micro_b001 and micro_b002 you can call:

env/bin/python src/my_scripts/files_to_pred_and_viz.py --out <folder> --model <path to pre-trained model> --text-ids micro_b001,micro_b002

In the output folder you will find 2 files for each text: one containing the predictions, and one being the graph image.

You can also use the segmentation obtained with the DiscourseSegmenter. Since getting such segmentations for micro-texts takes a while (mate tools are quite slow) I include here in data/corpus/german_mate_segmented all the texts already segmented. Same as before you can get predictions for these texts and visualize their graphs, with:

env/bin/python src/my_scripts/files_to_pred_and_viz.py --dir <input dir> --suffix <segmented> --out <folder> --model <path to pre-trained model> --text-ids micro_b001,micro_b002

The only difference is that you have to pass a folder containing all the segmented texts, in this case data/corpus/german_mate_segmented and the a suffix argument segmented (which is the default) to retrieve the ADUs created with the mechanism we taksed about.

It is possible to create visualization of argumentation structure of any texts. In order to do so it is needed first to obtain ADUs from a text. To do so (DiscourseSegmenter) will be used.

Please make sure you install DiscourseSegmenter in a different virtualenvironment. A quick way to do so is to use the script in shortcuts named create_dseg_venv. This will create a new virtual enviroment with discourse segmenter installed. Moreover, in order to use this segmentation tool you need to install as well the Mate tools models to provide the input in the requested format. To do so, please follow the instruction in mateparser/models dir.

With the scripy matepar_parser.sh (always in shortcuts) you can process all files inside a folder to get directly the output of DiscourseSegmenter. Finally by calling:

env/bin/python src/my_scripts/files_to_pred_and_viz.py --dir <dir with dseg files> --suffix <e.g. .seg> --model <trained evidencegraph model path> --out <out dir>

you can save the predictions on your texts and their visualization in a png file.

In order to provide a quick way to do this, on a single file it is possible to use:

user@machine:argumentation_minig ./shortcuts/text2arg.sh -i <input file> -o <out folder>

to directly go from a simple text to its argumentation structure visualization.

1. Joint prediction in MST-style discourse parsing for argumentation mining
   Andreas Peldszus, Manfred Stede.
   In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), Portugal, Lisbon, September 2015.

2. Mining Italian Short Argumentative Texts Ivan Namor, Pietro Totis, Samuele Garda, Manfred Stede. In: CLIC 2019