KTH Tangrams: A Dataset for Research on Alignment and Conceptual Pacts in Task-Oriented Dialogue

KTH Tangrams: A Dataset for Research on Alignment and Conceptual Pacts in Task-Oriented Dialogue Todd Shore, Theofronia Androulakaki, Gabriel Skantze KTH Speech, Music and Hearing Stockholm, Sweden tcshore@kth.se, androu@kth.se, gabriel@speech.kth.se Abstract There is a growing body of research focused on task-oriented instructor-manipulator dialogue, whereby one dialogue participant initiates a reference to an entity in a common environment while the other participant must resolve this reference in order to manipulate said entity. Many of these works are based on disparate if nevertheless similar datasets. This paper described an English corpus of referring expressions in relatively free, unrestricted dialogue with physical features generated in a simulation, which facilitate analysis of dialogic linguistic phenomena regarding alignment in the formation of referring expressions known as conceptual pacts. Keywords: reference, conceptual pacts, task-oriented dialogue 1. Introduction There is recent interest in the role of referring expres- sions (REs) in situated dialogue and the alignment of re- ferring language (RL) between dialogue participants (Barr and Keysar, 2002; Foster et al., 2006; Zarrieß et al., 2016; Aina et al., 2017). These datasets are useful for studying general patterns of alignment but are not specifically tai- lored to studying the effects of conceptual pacts (CPs) on RL in dialogue: CPs are patterns of RL which are mutu- ally accepted (either explicitly or implicitly) and used by all dialogue participants throughout the course of a dialogue (Brennan and Clark, 1996). In order to study this phenomenon, we introduce a collec- tion of recorded spoken English dialogues situated in a task called KTH Tangrams, wherein two participants collaborate in order to correctly select a predetermined abstract image on a procedurally-generated game board: Participants take turns assuming the role of either instructor, who can see which piece must be selected, or manipulator, who can select a piece but cannot see which one must be selected. This experiment design is similar to that used for many other works regarding RL, the most similar of these being PentoRef’s PentoCV and RDG-Pento (Zarrieß et al., 2016). PentoCV and RDG-Pento consist of one participant in- structing the other which pentomino piece (Golomb, 1994) is to be manipulated, but both participants are allowed to speak in a free fashion, a design originally defined by Kou- sidis et al. (2012). KTH Tangrams, however, is especially well-suited to observing CPs because the experiment de- sign entails participants deterministically referring to ab- stract entities multiple times in a dynamic environment without the entities themselves playing a role in a larger, culminating goal as done by e.g. Foster et al. (2006). 2. Related Work While there are many different works concerned with task- oriented dialogue, there are a number of differences in ex- periment design among them. 2.1. Static Versus Dynamic Environments The roles of instructor and manipulator seen in many tasks used for dialogue research are analogous to the roles of di- rector and matcher in traditional reference communication tasks, with the terms defined by Schober and Clark (1989) but the task itself originating from Krauss and Weinheimer (1964). …

Content truncated to 500 words