The corpus contains 136,996 (argumentative text, conclusion) pairs for the task of informative conclusion generation. For each argument in the corpus, argumentative knowledge such as discussion topic, conclusion targets and argument aspects are provided.
You can access the Webis-ConcluGen-21 corpus on Zenodo.
Given the vast amount of argumentative texts delivered by search engines, social media timelines, and similar, rendering their conclusions explicit would benefit accessability. Especially, these conclusions should be informative by emphasizing specific concepts from the argumentative text. Webis-ConcluGen-21 corpus is the first step towards developing models for informative conclusion generation. We extract argumentative knowledge (topic, targets, and aspects), encode them using control codes, and finetune a pretrained BART model as an abstractive approach for the task. Additionally, we employ a hybrid approach combining claim extraction and paraphrasing as a strong baseline.