Revealing and Predicting Online Persuasion Strategy with Elementary Units

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6274–6279, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6274 Revealing and Predicting Online Persuasion Strategy with Elementary Units Gaku Morio and Ryo Egawa and Katsuhide Fujita Tokyo University of Agriculture and Technology Koganei, Tokyo, Japan gakumorio@gmail.com, egawa@katfuji.lab.tuat.ac.jp, katfuji@cc.tuat.ac.jp Abstract In online arguments, identifying how users construct their arguments to persuade others is important in order to understand a persua- sive strategy directly. However, existing re- search lacks empirical investigations on highly semantic aspects of elementary units (EUs), such as propositions for a persuasive online argument. Therefore, this paper focuses on a pilot study, revealing a persuasion strategy us- ing EUs. Our contributions are as follows: (1) annotating five types of EUs in a persuasive forum, the so-called ChangeMyView, (2) re- vealing both intuitive and non-intuitive strate- gic insights for the persuasion by analyzing 4612 annotated EUs, and (3) proposing base- line neural models that identify the EU bound- ary and type. Our observations imply that EUs definitively characterize online persua- sion strategies. The annotated dataset, annotation guideline, and implementation of the neural model are available in public.1 1 Introduction Persuasive discourses are common in social me- dia. Recently, the automatic identification of such persuasive discourse is receiving attention in the field of computational linguistics (Habernal and Gurevych, 2016b; Tan et al., 2016; Habernal and Gurevych, 2016a; Persing and Ng, 2017; Musi et al., 2018; Hidey and McKeown, 2018; Ji et al., 2018; Durmus and Cardie, 2019; Gleize et al., 2019). Existing studies have examined persuasive arguments with regard to dynamics or lexical fea- tures (Tan et al., 2016), claim and premise analysis (Musi and Aakhus, 2018), and concessions (Musi 1Dataset and annotation guideline: http: //katfuji.lab.tuat.ac.jp/nlp_datasets/. Implementation: https://github.com/EdoFrank/ emnlp2019_morio_egawa !"#$%$&'()*)+'&%',')$-.$)*)./)0!1$-&'223)4563 !"##$%&'("$)*&)$+&,$&$(-)$*&.,*/$(0)$1$)*-23$1$&%$-2,-42-5-6&2)! "!"#$%#7$$8*"$02-9".,"$-,$%06*$(-44".$)*&2$&2')*-24$&$:".,;2$6&2$ -%&4-2"$"!"#$%#<$$1$=;2>)$)*-23$)*&)$&2')*-24$1$+-##$"9".$=;$+-##$ %&))".$("6&0,"$;5$)*-,$"!"#$%#7$$8*-,$:-6)0."$."&55-.%"=$*;+$1$5"#)$ ?*)):@AA-%40.76;%A9'B-CD07E$"!"#$%# 1$&%$2;2F."#-4-;0,$ "&%'()*+,-# &2=$1$40",,$)*&)$%-4*)$("$:&.)$;5$+*'$1$5""#$)*-,$+&'$ ?G;$:".,;2&#$4;=E$"!"#$%#77H$I&((."9-&)"=J$ BK$&0)*;. 1$=;2>)$32;+$+*;$';0$&."$&2=$+*&)$';0$*&9"$;.$*&9"2>)$=;2"$-2$ #-5"$"!"#$%#/$&2=$)*"."5;."$1$6&22;)$*;2",)#'$)"##$';0$+*&)$';0.$ -2=-9-=0&#$+;.)*$&,$&$:".,;2$%&'$;.$%&'$2;)$("!"!"#$%#7$1>%$L0,)$ ,;%"$40'$;2$)*"$-2)".2")$"!"#$%#7$M0)H$I&((."9-&)"=J$H ! $%&'(')* N;0.$-%40. #-23$=;",2>)$,""%$);$+;.37 !*"2$';0$)&#3$&(;0)$)*-24,$';0$=;$%&))".-24/$"O&6)#'$*;+$=;$ ';0$%"&2$-)P$"./%(+0)1"#23("(%*%,(# Q;+$(-4$;5$&$6;2).-(0)-;2$=;$';0$*&9"$);$%&3"$("5;."$';0$="6-="$ )*&)$-)$%&))".,P$"./%(+0)1"#23("(%*%,(# H$I&((."9-&)"=J$H +*,-(')* R+&.=$ &$="#)& Figure 1: Overview of our EU annotation in Change- MyView. We observed that Testimonies tend to appear in an original post (OP) (see Section 4.) et al., 2018). However, to the best of our knowl- edge, fine-grained analyses with an elementary unit (EU) semantic, such as propositions, are still in their infancy, although we estimate that EUs are important in obtaining the persuasion strategy. Also, existing studies do not extract token-level EUs via machine learning, which would be a basic step for an automatic understanding of the strat- egy. Therefore, this paper shows the fundamental role of EUs in a persuasive forum by annotat- ing five types of token-level EUs (i.e., Testimony, Fact, Value, Policy, and Rhetorical Statement) in a persuasive forum and also provides a baseline neural model to identify the EUs automatically.2 In this study, we use the effective dataset of ChangeMyView (Tan et al., 2016), which is a subreddit where original post (OP) users provide a controversial view in the title to change their 2Although …

Content truncated to 500 words