A multi-task approach to argument frame classification at variable granularity levels

Abstract

Within the field of argument mining, an important task consists in predicting the frame of an argument, that is, making explicit the aspects of a controversial discussion that the argument emphasizes and which narrative it constructs. Many approaches so far have adopted the framing classification proposed by Boydstun et al. [3], consisting of 15 categories that have been mainly designed to capture frames in media coverage of political articles. In addition to being quite coarse-grained, these categories are limited in terms of their coverage of the breadth of discussion topics that people debate. Other approaches have proposed to rely on issue-specific and subjective (argumentation) frames indicated by users via labels in debating portals. These labels are overly specific and do often not generalize across topics. We present an approach to bridge between coarse-grained and issue-specific inventories for classifying argumentation frames and propose a supervised approach to classifying frames of arguments at a variable level of granularity by clustering issue-specific, user-provided labels into frame clusters and predicting the frame cluster that an argument evokes. We demonstrate how the approach supports the prediction of frames for varying numbers of clusters. We combine the two tasks, frame prediction with respect to media frames categories as well as prediction of clusters of user-provided labels, in a multi-task setting, learning a classifier that performs the two tasks. As main result, we show that this multi-task setting improves the classification on the single tasks, the media frames classification by up to +9.9 % accuracy and the cluster prediction by up to +8 % accuracy.

Frame-oriented Summarization of Argumentative Discussions

“…In cases where the model has lower confidence in its prediction, the argument may consist of multiple frames. This overcomes the limitation of clustering-based approaches and classifiers which strictly assign a single frame to arguments that may contain multiple ones Heinisch and Cimiano, 2021).…”

“…Extending the state-of-theart frame classification model of Heinisch and Cimiano (2021), we developed a new classifier trained on an external frame-labeled dataset. The existing classifier of Heinisch and Cimiano (2021), utilizes a recurrent neural network to assign a single frame to an argument, and combines it with a model that predicts a cluster of frame labels from the inventory of Ajjour et al (2019) in a multi-task setting. Particularly longer arguments, however, often contain multiple frames.…”

“…Later, Ajjour et al (2019) modeled framing in argumentation more systematically, introducing automatically extracted, fine-grained, issue-specific frame labels. Heinisch and Cimiano (2021) successfully combined computational argumentation with framing theory by showing a latent connection between the different frame granularities for the media frames defined by Boydstun et al (2014). Hartmann et al (2019) also used frame-labeled data from newswire corpus to successfully train frame classifiers for political discussions via multi-task and adversarial learning.…”

Key Point Analysis via Contrastive Learning and Extractive Argument Summarization

“…We employ the tagger of Schiller et al (2021) to extract the arguments' aspects (on average, 2.1 aspects per argument). To tackle the lack of diversity, we follow Heinisch and Cimiano (2021) final set of key points summarizing the arguments for a given topic and stance maximizes the coverage of the set of arguments' aspects. To this end, we apply greedy approximation for selecting our candidates, where an argument sentence is chosen if it covers the maximum number of unique aspect clusters while having the smallest overlap with the clusters covered by the already selected candidates.…”

Key Point Analysis via Contrastive Learning and Extractive Argument Summarization

“…We employ the tagger of Schiller et al (2021) to extract the arguments' aspects (on average, 2.1 aspects per argument). To tackle the lack of diversity, we follow Heinisch and Cimiano (2021) final set of key points summarizing the arguments for a given topic and stance maximizes the coverage of the set of arguments' aspects. To this end, we apply greedy approximation for selecting our candidates, where an argument sentence is chosen if it covers the maximum number of unique aspect clusters while having the smallest overlap with the clusters covered by the already selected candidates.…”