T5-frame-identification

Code for task of course 'deep learning for natural language processing', at University Bielefeld. The T5 transformer is finetuned for the Task of frame-identification.

Find a description of the task and code here

The scripts should automatically use your GPU if CUDA is correctly installed and works with tensorflow2. Training and evaluation is extremely slow on CPU! Testing is possible. All code was tested using python3.8.9

To use the code I suggest to use a virtuelenv (e.g. pip or conda) and install the necessary packages. This can be done by calling python setup.py from the root of this repository. The Webis-Argument-Framing-19 dataset and the trained T5 will also be downloaded and extracted to the correct location. Should this fail you can place the files manually:

• Download the dataset. The 'Webis-argument-framing.csv' needs to be stored in a subfolder called 'data'.
• The python packages can be installed by calling python -m pip install -r requirements.txt from the root of this repository.
• Download the trained model here and extract the archive in this folder.

In addition to the setup.py two executable scripts are provided. To train and evaluate the t5 call python train_evaluate_main.py. To test if the code is working supply the --debug option. This is recommended when using CPU. To fetch some random examples from the dataset and run a pretrained model on them call e.g. python test.py --test-no 3. The --test-no option can be used to print multiple examples in a single run.

All scripts use reasonable defaults defined in defaults.py. If you want to tweak them you can do so. It is also possible to call the scripts with options. Call them with the -h option to print information on what options are available.

The best pretrained model will be in trained_models/t5-base-frame-finetuned. All newly trained models will be stored in the same location by default. If you want to train another model be sure to supply a new name for it via --model-savename option or set another model_savename in the defaults.py. Setting the name in defaults.py makes it easy to call test.py afterwards.

The Layout of this folder after calling setup.py and extracting the model should be:

.
+-- data
|   +-- Webis-argument-framing.csv
|   +-- args_test.pkl
+-- defaults.py
+-- download_model.sh
+-- evaluate.py
+-- model.py
+-- preprocessing.py
+-- README.md
+-- requirements.txt
+-- setup.py
+-- test.py
+-- trained_models
|   +-- t5-base-frame-finetuned
    |   +-- config.json
    |   +-- tf_model.h5
+-- train-evaluate-main.py
+-- utils.py