Webis-ConcluGen-2021

Update: Patch 1.0.1 removes duplicates (up to 5% of the examples in the base version) and re-indexes the examples with new "id" values. The argument and conclusion are now merged into a single csv file for each split to be easily loaded as dataframes. We also include the test set used for manual evaluation (manual_evaluation_arguments.csv) to support qualitative comparisons. 

The corpus contains 130.519 (argumentative text, conclusion) pairs. There are four variations of this corpus:

1. base:  (argumentative text, conclusion) pairs
2. topic: (topic encoded argumentative text, conclusion) pairs
3. targets: (topic and target encoded argumentative text, conclusion) pairs
4. aspects: (topic and aspect encoded argumentative text, conclusion) pairs

Each variation is split into train, val, and test files. Additionally, the test set used for automatic evaluation is provided (automatic_evaluation_test_set.csv) which must be used for any quantitative evaluation.