End-to-end Argument Mining with Cross-corpora Multi-task Learning

End-to-end Argument Mining with Cross-corpora Multi-task Learning Gaku Morio, Hiroaki Ozaki, Terufumi Morishita, and Kohsuke Yanai Research and Development Group Hitachi, Ltd. Kokubunji, Tokyo, Japan {gaku.morio.vn,hiroaki.ozaki.yu, terufumi.morishita.wp,kohsuke.yanai.cs}@hitachi.com Abstract Mining an argument structure from text is an important step for tasks such as argu- ment search and summarization. While studies on argument(ation) mining have proposed promising neural network models, they usu- ally suffer from a shortage of training data. To address this issue, we expand the training data with various auxiliary argument mining cor- pora and propose an end-to-end cross-corpus training method called Multi-Task Argument Mining (MT-AM). To evaluate our approach, we conducted experiments for the main argu- ment mining tasks on several well-established argument mining corpora. The results demon- strate that MT-AM generally outperformed the models trained on a single corpus. Also, the smaller the target corpus was, the better the MT-AM performed. Our extensive analyses suggest that the improvement of MT-AM depends on several factors of transferability among auxiliary and target corpora. 1 Introduction Argument(ation) mining (AM), the task of iden- tifying an argument structure from text, has been gaining attention in recent years (Stede and Schneider, 2018; Lawrence and Reed, 2019). Also known as argument(ation) structure parsing (Kuribayashi et al., 2019), AM typically iden- tifies argumentative component spans, classifies the type of the components, and classifies the relations between the components. In span identi- fication, we discriminate argumentative text units (i.e., spans) from non-argumentative ones in a given text. In component classification, we clas- sify the spans into argumentative labels such as Claim and Premise. Relation classification detects argumentative links between the components and classifies each link into a relation label such as Support and Attack. Researchers have been utilizing various datasets (AM corpora) to develop AM models. These corpora are designed on the basis of differ- ent theoretical frameworks and conceptualizations (Daxenberger et al., 2017). For example, Peldszus and Stede (2016) developed a corpus called the Microtext Corpus (MTC) on the basis of Freeman’s theory of the macro-structure of arguments (Freeman, 2011) to capture hypothetical dialecti- cal exchanges. Stab and Gurevych (2017) created the Argument-annotated Essays Corpus (AAEC), in which a connected tree is used to represent the structure of each argument in a paragraph. A notable feature of these AM corpora is that, due to the lack of a clear underlying framework on which to base the annotation, new annotated corpora are likely to diverge in terms of the span, type of components, and type of relations. In addition, developing an AM corpus is ex- pensive because of the annotation cost (Schulz et al., 2018). For example, Lauscher et al. (2018) conducted multiple calibration phases when train- ing annotators to improve the low inter-annotator agreement, which drove up costs. This suggests that it would be infeasible to create a sufficiently large AM corpora for a new domain, and the resultant lack of data inevitably degrades the performance of argument structure parsing. Given the lack of a large annotated AM corpus, multi-task learning on AM corpora la- …

Content truncated to 500 words