baseline training on `YA-I2L` data

[ArneBinder]

This adds configs to train a simple baseline model on the YA-I2L data from the DialAM-2024 dataset. Please have a look into the relevant configs (in the respective config subfolders):

• dataset: dialam2024_base (pure pie-document dataset) and dialam2024_prepared (+validation split)
• model: sequence_classification_with_pooler (the trainable Pytorch-Lightning model), see pie-modules for further information
• taskmodule: re_text_classification_with_indices (conversion of documents to model inputs and back), see pie-modules for further information
• experiment: dialam2024_ya_i2l (sticking everything together)

Approach: Since we pre-calculate the alignment of I- and L-nodes, we can frame the task of YA-I2L-relation extraction as unary-relation extraction where each L-node participates in exactly one relation.

Execute a fast dev run (one batch only):

python src/train.py \
experiment=dialam2024_ya_i2l \
+trainer.fast_dev_run=true

train on the GPU:

python src/train.py \
experiment=dialam2024_ya_i2l \
trainer=gpu

Notes:

• This requires #17.
• This updates the dependencies, you may need to uninstall pytorch-ie, pie-modules, and pie-datasets and reinstall via pip install -r requirements.txt

[tanikina]

We used this code to train our first model for YA-I2L relations with bert-base-uncased (default configuration, only the batch size was reduced from 32 to 8).

Here are the results on the validation set with micro-F1 0.94 and macro-F1 0.34 (note that some classes occur only rarely in the validation split, e.g., Agreeing-1 or Challenging-2):

Results for the classes that appear more than 10 times in the validation set:	class	support	f1
Asserting	1704	0.976
AssertiveQuestioning	19	0.294
NONE	39	0.260
PureQuestioning	111	0.744
RhetoricalQuestioning	18	0.450

W&B run

TODO: in logs we are getting warnings that need to be checked (probably related to offset computation):

[pie_modules.taskmodules.re_text_classification_with_indices][WARNING] - doc.id=25497: Skipping invalid example, cannot get argument token slices for {​LabeledSpan(start=51, end=185, label='L', score=1.0): "Michelle O'Neill : I would remind viewers that the majority of people and elected representatives in the north are opposed to Brexit.\xa0"}​