1 - 2 of 2
- PublicationThe Dimensions and Adaptation of Partner Models in Human-Machine Dialogue(University College Dublin. School of Information and Communication Studies, 2022)While speech interfaces have been widely adopted in recent years it appears people tend to use them in a limited fashion, rarely exploring their full functionality. Research suggests this is largely due to systems being perceived as having limited communicative abilities. Yet, progress in developing a nuanced understanding of user perceptions in human-machine dialogue (HMD) interaction have been hampered by a lack of theory building and measure development. This gap is addressed here through a series of studies aimed at developing a theoretical concept known as partner modelling. A partner model is a cognitive representation of a dialogue partner that reflects their perceived communicative ability and social relevance. Research suggests partner models influence interaction behaviour in both human-machine and human-human dialogue (HHD). However, the concept is currently under-defined and theoretical accounts contain a number of untested assumptions, including: the idea that people adapt their partner models to reflect interaction experiences, that this adaptation requires substantial cognitive resources, and that adaptation may be reflected in a language phenomena known as lexical alignment. Work presented here investigates the dimensionality and adaptation of partner models in the context of HMD interaction. Study 1 uses a mixed-method approach to explore the breadth of terms people use to define the communicative ability of speech interface technologies. In Study 2 this data was complimented by a review of relevant literature and used to develop a standardized subjective measure of partner models in the context of HMD. The scale highlights three salient dimensions of partner models in HMD: perceived communicative competence and dependability; perceived human-likeness in communication; and perceived communicative flexibility. Study 3 then looked to validate its factor structure, convergent and divergent validity, and test-retest reliability. Finally, Study 4 uses the measure to examine the impact of interaction experience (i.e., system errors) on partner models, lexical alignment and cognitive load. Findings suggest experiencing system errors in interaction can provoke adaptation of partner models, that this does not require substantial cognitive resources and that it is not reflected in lexical alignment behaviour. The implications of these findings and limitations of the work are then discussed in the closing sections.
- PublicationHow Early Rewards Influence Choice: Targeting model-free processing through reward timing(University College Dublin. School of Information and Communication Studies, 2021)While many people claim to have the intention to perform certain behaviours, it is commonly the case these intentions do not come to fruition. This issue is particularly pronounced in cases where there is a long delay between intention and the behaviour, or cases where there is a strong automatic impulse that acts against the intention. According to dual-process theories, this intention-behaviour gap is a result of a conflict between two types of systems: a habitual model-free system and a deliberate model-based system. Usually, interventions target the model-based system, providing important information necessary to convince individuals that the behaviour is desirable or beneficial. However, this approach mostly ignores the model-free system, leaving a large part of the decision-making process outside of the intervention. The early reward strategy is a method to target the model-free system directly and considers the known mechanisms behind how reward information is processed. In particular, it focuses on how reward timing affects decision making within a sequence of actions. Due to how temporal discounting and temporal difference learning lead to reductions in the value of the reward based on how far it is placed from the first action in the sequence, placing the reward as close to the start of the sequence as possible is likely to prevent this reduction from occurring as much as possible. This early reward strategy was tested across four experiments and was found to successfully alter behaviour in a way predicted by the theory. Two of the experiments focused on a computational approach, using reinforcement learning algorithms to predict behaviour and compare it against the participant responses. The other two experiments were conducted with a more applied approach that used tasks more representative of real-world action sequences to test the extent to which behaviour was affected by early rewards. Whether the reward was monetary or gamified, placing a reward earlier in a sequence improved the frequency of selection for that sequence significantly when compared to other reward placements. The results have important implications for anyone attempting to incentivise new behaviours by providing a theory-driven approach towards maximising the effectiveness of the reward, particularly to the model-free system. As a result, consideration for reward timing should be integral to any incentive system that involves sequences of actions, with a strong emphasis on providing rewards as early in the interaction as possible.