Creating Large-Scale Argumentation Structures for Dialogue Systems

Creating Large-Scale Argumentation Structures for Dialogue Systems Kazuki Sakai1, Akari Inago2, Ryuichiro Higashinaka3, Yuichiro Yoshikawa1, Hiroshi Ishiguro1, Junji Tomita3 1Osaka University, 2Ochanomizu University, and 3NTT Media Intelligence Laboratories Abstract We are planning to develop argumentative dialogue systems that can discuss various topics with people by using large-scale argumenta- tion structures. In this paper, we describe the creation process of these argumentation structures. We created ten structures each having more than 2000 nodes of five topics in English and five topics in Japanese. We analyzed the created structures for their characteristics and investigated the differences between the two languages. We conducted an evaluation experiment to ascertain that the structures can be applied to dialogue systems. We conducted another experiment to use the created argumentation structures as training data for augmenting the current argumentation structures. Keywords: Argumentation, Large-scale argumentation structures, Dialogue systems 1. Introduction Argumentation is a process of reaching consensus through premises and rebuttals and is important for making decisions and exchanging views. Argumen- tation has long been studied in the fields of rhetoric, informal logic, and more recently artificial intelli- gence. For example, there have been studies on automatically extracting conclusions and premises from documents (Rosenthal and McKeown, 2012; Yanai et al., 2016; Lippi and Torroni, 2016). Other studies have devised argumentation models (Toulmin, 1958; Reed and Rowe, 2004; Walton, 2013) and visualizations of the models (Gordon et al., 2007; Reed and Rowe, 2004; Snaith et al., 2010). To develop dialogue systems that can support humans in argumentation, we are planning to develop argumentative dialogue systems that can discuss various topics with people by using large-scale argumentation structures; we argue that large-scale argu- mentation structures are necessary for systems to respond appropriately to various arguments raised by users. Recently, corpora containing argumentation struc- tures of discussions/meetings have been made avail- able (Janin et al., 2003; Renals et al., 2007). Other studies have extracted argumentation structures from corpora (Fern´andez et al., 2008; Bui et al., 2009). However, these structures are small; thus, insuffi- cient as knowledge for dialogue systems. One of the largest argumentation databases currently available is AIFdb (Lawrence et al., 2012), which is an open database containing argumentation structures in argument inter- change format (AIF). Although AIFdb contains many argumentation structures, each structure is small; for exam- ple, the largest contains about 250 nodes, and the average number of nodes per structure is 8.14 (See Table 1). In this paper, we describe the creation process to con- struct large-scale argumentation structures for dialogue sys- tems. We manually created several large-scale argumen- tation structures based on a conventional argumentation model (Walton, 2013), and each structure has more than 2000 nodes. We created the structures in two languages; English (major language) and Japanese (author language). To verify the effectiveness of the created argumentation Figure 1: Argumentation model structures, we conducted an evaluation experiment to as- certain that the structures can be applied to dialogue sys- tems. We conducted another experiment to use the created argumentation structures as training data for augmenting …

Content truncated to 500 words