Spectral Analysis of Information Density in Dialogue Predicts Collaborative Task Performance

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 623–633 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1058 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 623–633 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1058 Spectral Analysis of Information Density in Dialogue Predicts Collaborative Task Performance Yang Xu and David Reitter College of Information Sciences and Technology The Pennsylvania State University University Park, PA 16802, USA yang.xu@psu.edu, reitter@psu.edu Abstract We propose a perspective on dialogue that focuses on relative information contribu- tions of conversation partners as a key to successful communication. We predict the success of collaborative task in English and Danish corpora of task-oriented dia- logue. Two features are extracted from the frequency domain representations of the lexical entropy series of each interlocutor, power spectrum overlap (PSO) and rela- tive phase (RP). We find that PSO is a neg- ative predictor of task success, while RP is a positive one. An SVM with these fea- tures significantly improved on previous task success prediction models. Our find- ings suggest that the strategic distribution of information density between interlocu- tors is relevant to task success. 1 Introduction What factors affect whether information is con- veyed effectively and reliably in conversations? Several theoretical frameworks have emerged that model dialogical behavior at different granularity levels. Can we use them to measure communica- tive effectiveness? Grounding theory (Clark and Brennan, 1991) models a successful communication as a pro- cess during which “common ground” (i.e., mu- tual knowledge, beliefs etc.) is jointly built among interlocutors. The interactive alignment model (IAM) (Pickering and Garrod, 2004) proposes that the ultimate goal of dialogue is the alignment of interlocutors’ situational model, which is helped by alignment at all other lower representation lev- els (e.g., lexical, syntactic etc.), driven by the psy- chologically well-documented priming effects. Recently, empirical studies have verified the explanatory powers of the above-mentioned the- ories, especially the IAM, utilizing dialogues recorded and transcribed from various collabora- tive tasks conducted in laboratory settings (Reit- ter and Moore, 2007; Reitter and Moore, 2014; Fusaroli et al., 2012; Fusaroli and Tyl´en, 2016). In those studies, the quality of communication is directly reflected in the collaborative performance of interlocutors, i.e., how successful they are in ac- complishing the task. Although they do not come to fully agree on which theoretical accounts of di- alogue (e.g., interactive alignment vs. interper- sonal synergy) provides better explanations (see Section 2.1 for details), the majority of these stud- ies have confirmed that the alignment of certain linguistic markers, lexical items, or syntactic rules between interlocutors correlates with task success. What is missing from the picture, however, is the computational understanding of how strategies of interaction and the mix of information contribu- tions to the conversation facilitate successful com- munication. This is understandable because those higher level concepts do not directly map onto the atomic linguistic elements and thus are much more difficult …

Content truncated to 500 words