Efficient Argument Structure Extraction with Transfer Learning and Active Learning

Findings of the Association for Computational Linguistics: ACL 2022, pages 423 - 437 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Efficient Argument Structure Extraction with Transfer Learning and Active Learning Xinyu Hua Bloomberg New York, NY xhua22@bloomberg.net Lu Wang Computer Science and Engineering University of Michigan Ann Arbor, MI wangluxy@umich.edu Abstract The automation of extracting argument struc- tures faces a pair of challenges on (1) en- coding long-term contexts to facilitate com- prehensive understanding, and (2) improving data efficiency since constructing high-quality argument structures is time-consuming. In this work, we propose a novel context-aware Transformer-based argument structure predic- tion model which, on five different domains, significantly outperforms models that rely on features or only encode limited contexts. To tackle the difficulty of data annotation, we ex- amine two complementary methods: (i) trans- fer learning to leverage existing annotated data to boost model performance in a new target do- main, and (ii) active learning to strategically identify a small amount of samples for anno- tation. We further propose model-independent sample acquisition strategies, which can be generalized to diverse domains. With exten- sive experiments, we show that our simple-yet- effective acquisition strategies yield competi- tive results against three strong comparisons. Combined with transfer learning, substantial F1 score boost (5-25) can be further achieved during the early iterations of active learning across domains. 1 Introduction Identifying and understanding the argumentative discourse structure in text has been a critical task in argument mining (Peldszus and Stede, 2013; Cabrio and Villata, 2018; Lawrence and Reed, 2019; Li et al., 2020). It plays an important role of discovering the central theses and reasoning process across a wide spectrum of domains, from formal text such as legal documents (Palau and Moens, 2009; Lippi and Torroni, 2016; Poudyal et al., 2020) and scientific literature (Mayer et al., 2020; Fergadis et al., 2021; Al Khatib et al., 2021), to online posts and discussions (Cardie et al., 2008; Boltuži´c and Šnajder, 2014; Park and Cardie, 2014; Habernal and Gurevych, 2017; Hua and Wang, Peer Review eRulemaking Online Comment (1) I think this submission does not meet the community standard. (2) The originality of the approach is unclear. (3) Most existing work learning embeddings of multi-relational graphs also create multiple examples. (4) The difference is that here it is done sequentially, not really meaningful. (5) Secondly, none of the baselines in the experiments use more information than GEN (the original graph), (6) which is unfair comparison. (1) I don’t think that the partial SSN should be included. (2) That raises too many privacy concerns. (3) And I’m sure that one day a debt collection company will have a “system malfunction” and will “accidentally” send the full SSN. (4) Lastly, many companies classify a partial SSN as the first 5 digits whereas other stick to the traditional last 4 digits. (5) If a mail thief is lurking, then he might have access to a consumers full SSN Figure 1: Excerpts of arguments in peer reviews and online comments. On the right, …

Content truncated to 500 words