Adjacency Pair Recognition in Wikipedia Discussions using Lexical Pairs

PACLIC 28 !479 Adjacency Pair Recognition in Wikipedia Discussions using Lexical Pairs Emily K. Jamison ‡ and Iryna Gurevych †‡ ‡Ubiquitous Knowledge Processing Lab (UKP-TUDA), Department of Computer Science, Technische Universit¨at Darmstadt † Ubiquitous Knowledge Processing Lab (UKP-DIPF), German Institute for Educational Research http://www.ukp.tu-darmstadt.de Abstract Adjacency pair recognition, a necessary com- ponent of discussion thread reconstruction, is the task of recognizing reply-to relations be- tween pairs of discussion turns. Previously, dialogue act classification and metadata-based features have been shown useful in adjacency pair recognition. However, for certain forums such as Wikipedia discussions, metadata is not available, and existing dialogue act typologies are inapplicable. In this work, we show that adjacency pair recognition can be performed using lexical pair features, without a dialogue act typology or metadata, and that this is ro- bust to controlling for topic bias of the discus- sions. 1 Introduction A growing cache of online information is contained inside user-posted forum discussions. Thread struc- ture of the discussion is useful in extracting infor- mation from threads: Wang et al. (2013) use thread structure to improve IR over threads, and Cong et al. (2008) use thread structure to extract question- answer pairs from forums. However, as Seo et al. (2009) point out, thread structure is unavailable in many forums, partly due to the popularity of fo- rum software phpBB1 and vBulletin2, whose default view is non-threaded. Thread reconstruction provides thread structure to forum discussions whose original thread struc- ture is nonexistant or malformed, by sorting and re- ordering turns into a directed graph of adjacency (reply-to) relations. Pairs of adjacent turns (adja- cency pairs) were first identified by Sacks et al. 1http://www.phpbb.com/ 2http://www.vbulletin.com/ Turn1: This article has been gutted. I deleted a lot of the cruft that had taken over, but a lot of former material is missing.[...] Turn2: Good; the further this nest of doctrinaire obscuri- ties is gutted, the better. Turn3: Wait, you changed it to say that English doesn’t have a future tense or you’re citing that as an error (which it would naturally be)? For what it matters, [...] Turn4: English doesn’t have a future tense. It indicates the future with a modal (will) used with the present-tense inflection of the verb. [...] Figure 1: Excerpt from the EWDC discussion Grammatical Tense:gutted. (1974) as the structural foundation of a discussion, and recognition of adjacency pairs is a critical step in thread reconstruction (Balali et al., 2014; Wang et al., 2008; Aumayr et al., 2011). Figure 1 shows an excerpt from Ferschke’s (2014) English Wikipedia Discussions Corpus. Thread structure is indicated by tab indents. Turn pairs (1,2), (1,3), and (3,4) are adjacency pairs; pairs (2,3) and (1,4) are not. Adjacency pair recognition is the classification of a pair of turns as adjacent or nonad- jacent. Although most previous work on thread recon- struction takes advantage of metadata such as user id, timestamp, and quoted material (Aumayr et al., 2011; Wang et al., 2011a), metadata is unreliable in some forums, such as Wikipedia Discussion page forums, …

Content truncated to 500 words