Fine and Coarse Granular Argument Classification before Clustering

Computational argumentation and especially argument mining together with retrieval enjoys increasing popularity. In contrast to standard search engines that focus on finding documents relevant to a query, argument retrieval aims at finding the best supporting and attacking premises given a query claim, e.g., from a predefined collection of arguments. Here, a claim is the central part of an argument representing the standpoint of a speaker with the goal to persuade the audience, and a premise serves as evidence to the claim. In addition to the actual retrieval process, existing work has focused on (1) classifying polarities of arguments into supporting or opposing, (2) classifying arguments by their frames (such as economic or environmental), and (3) clustering similar arguments by their meaning to avoid repetitions in the result list. For experiments, either hand-made argument collections or arguments extracted from debate portals were used. In this paper, we extend existing work on argument clustering, making the following contributions: First, we introduce a novel pipeline for clustering arguments. While previous work classified arguments either by polarity, frame, or meaning, our pipeline incorporates these three, allowing a more systematic presentation of arguments. Second, we introduce a new dataset consisting of 365 argument graphs accompanying more than 11,000 high-quality arguments that, contrary to previous datasets, have been generated, displayed, and verified by journalists and were published in newspapers. A thorough evaluation with this dataset provides a first baseline for future work.