Multi-Task Attentive Residual Networks for Argument Mining

Abstract

We explore the use of residual networks and neural attention for multiple argument mining tasks. We propose a residual architecture that exploits attention, multi-task learning, and makes use of ensemble, without any assumption on document or argument structure. We present an extensive experimental evaluation on five different corpora of user-generated comments, scientific publications, and persuasive essays. Our results show that our approach is a strong competitor against state-of-the-art architectures with a higher computational footprint or corpus-specific design, representing an interesting compromise between generality, performance accuracy and reduced model size.

Argument mining as a multi-hop generative machine reading comprehension task

“…There are two ways to combine the remaining subtasks, i.e., ACC + ARIC and ACC + ARI + ARC. Considering that there is no established preference in existing literature-the former is used on the AbstRCT dataset (Mayer et al, 2020;Galassi et al, 2023) and the latter <non_AC> We assessed bleb morphology and the intraocular pressure (IOP)-lowering effect of trabeculectomy with ologen compared to mitomycin C (MMC) in juvenile open-angle glaucoma (JOAG)... <AC1>The mean postoperative IOP was statistically significantly lower than the mean preoperative IOP at each follow-up in each group. <AC2> At 1 year, the mean postoperative IOP was significantly lower in group A.…”

“…For example, some models only leverage single ACs or a pair of ACs as input without wider contextual information (Mayer et al, 2020;Accuosto et al, 2021). Despite using similar input, Galassi et al (2023) recognised that individual propositions may not always provide enough information for AM.…”

“…Further, some researchers found that combining plain text with its structural information is beneficial for AM. Examples include the distance between two ACs (Galassi et al, 2023); the position of the AC in the document or paragraph (Potash et al, 2017;Bao et al, 2021); discourse parser features (Hewett et al, 2019); and argumentative zoning information . Although these types of information are helpful for the model to understand the structure of an argumentative text, a model cannot understand the "chain of thought" from such structural information and the plain text only.…”

“…Here, each AC is associated with at most one outgoing AR, which means that the argumentative graph of each paragraph has a tree or forest structure. For the AbstRCT dataset, we follow previous studies(Mayer et al, 2020;Galassi et al, 2023) by merging major claim and claim into a single category. All results on both datasets are averaged scores of three different random seeds and are reported as macro-averaged F1 scores.…”