Comparison of Common Ground Models for Human--Computer Dialogue

Abstract

Common ground processes [26] can improve performance in communication tasks [72, 42, 43, 24], and understanding these processes will likely benefit human--computer dialogue interfaces. However, there are multiple proposed theories with different implications for interface design. Fusaroli and Tylén [40] achieved a direct comparison by designing two models: one based on alignment theory and the other based on complementarity theory that encapsulated interpersonal synergy and audience design. The current research used these models, extending them to differentiate between interpersonal synergy and audience design. Few studies have tested multiple common ground models against tasks representative of envisioned human--computer interaction (HCI) applications. We report on four such tests, which allowed examination of generalizability of findings. Results supported the complementarity models over the alignment model, and were suggestive of the audience design variant of complementarity, providing guidance for HCI design that differs from contemporary approaches.

Audience Design and Egocentrism in Reference Production During Human-Computer Dialogue

“…As highlighted in Section 2.3, previous work has emphasised the importance of audience design in HCD language production (Branigan et al, 2011;Oviatt et al, 1998, Rothwell et al, 2021. This aligns with HHD work (Brennan et al, 2010;Clark, 1992Clark, ,1996Clark & Marshall, 1981) that declares dialogue to be a collaborative process where interlocutors adapt their language choices to be more felicitous to the needs of partners in conversation.…”

“…Yet this knowledge is critical for the ongoing efforts to computationally model user language behaviours in human-computer dialogue within the HCI field (Rothwell et al, 2021), whilst also being important so as to inform technological development and design. The current consensus is that users adapt their language choices based on the perception of the capabilities of computers as dialogue partners (Amalberti et al, 1993;Brennan, 1998;Cowan et al, 2019a;Le Bigot et al, 2007;Luger & Sellen, 2016;Meddeb Frenz-Belkin, 2010a;Rothwell et al, 2021), which can be influenced by design (Cowan et al, 2019), echoing the concept of audience design within human-human dialogue research (Bell, 1984). However, experimental findings comparing language choices with human and computer interlocutors have sometimes found little presence of adaptation and audience design (Cowan & Branigan, 2015;Cowan et al, 2015), questioning the ubiquity of such an account.…”

“…When talking to computers, participants focus more on the geometric features of the images, whereas when interacting with people, participants focus more on the image as a whole. Indeed, recent modelling of alternative explanations for language production in HCD gives strong evidence of audience design being a major driver of language production over other more mechanistic processes such as priming (Rothwell et al, 2021). This reliance on audience design is thought to be driven by users seeing computers as at-risk listeners (Oviatt et al, 1998) or basic dialogue partners (Branigan et al, 2011), leading them to adapt their language to increase the likelihood of communication success based on these perceptions.…”

“…Such work is important to inform current efforts to computationally model user language production in HCD (e.g. Rothwell et al, 2021), whilst also highlighting key theoretical issues (e.g. audience design) that may influence and be influenced by speech interface design (Cowan et al, 2019).…”

“…Despite this rapid growth, we know little about the psychological and linguistic mechanisms that people employ in speech-based human-computer dialogue (HCD). This knowledge is needed to help inform theory building within HCI on what influences user language interactions (Cowan et al, 2023;Peña et al, In press;Shen and Wang, 2023) supporting recent efforts for more formal and computation models of user behaviour in human-machine dialogue (e.g Rothwell et al, 2021). Such insights could also support speech interface development through informing speech technology components (e.g., dialogue management and recognition), whilst informing design decisions so as to take into account their influence on mechanisms for linguistic behaviour (Braunger et al, 2017, Zhao et al, 2022.…”