README.md

Content

This folder provides a configuration for running the SciBERT-CLS-modal model on the standard train/test split of MiST. Configurations are stored as JSON files. To enable re-use of configuration parts without duplication, configurations are organized in a modular way:

• cross_validation: Contains paths to train and validation files for each split in cross-validated training.
• evaluation: Configs for training and evaluating models in cross-validated training, i.e., training and evaluating a model on each of the splits. Can be used as an argument to evaluate.py.
• model: Configs for training and evaluating a single model on a given train-val-test split.
• task: Configs for tasks, e.g., classification on MiST. These configs are referred to in the tasks attribute of the model configs. Note that the identifier used for a specific task in tasks and heads of a model config (the latter for specifying task-specific output heads) and in cross-validation configs must be identical (e.g., mist).

We also provide an example configuration for multi-task training with EPOS (Marasovic et al. (2016). In order to run this configuration, you need to get the corpus and convert it to the following format:

SINGLE-LABEL

#<sentence ID>
<target modal verb, e.g., "can">
<position of target modal verb, 0-indexed>
<label of target modal verb, e.g., "dy">
<token_1>
<token_2>
...
<token_n>

#<sentence ID>
<target modal verb, e.g., "can">
<position of target modal verb, 0-indexed>
<label of target modal verb, e.g., "dy">
<token_1>
<token_2>
...
<token_m>