Evaluating humanoid embodied conversational agents in mobile guide applications

Doumanis, Ioannis (2013) Evaluating humanoid embodied conversational agents in mobile guide applications. PhD thesis, Middlesex University. [Thesis]

PDF - Final accepted version (with author's formatting)
Download (10MB) | Preview


Evolution in the area of mobile computing has been phenomenal in the last few years. The exploding increase in hardware power has enabled multimodal mobile interfaces to be developed. These interfaces differ from the traditional graphical user interface (GUI), in that they enable a more “natural” communication with mobile devices, through the use of multiple communication channels (e.g., multi-touch, speech recognition, etc.). As a result, a new generation of applications has emerged that provide human-like assistance in the user interface (e.g., the Siri conversational assistant (Siri Inc., visited 2010)). These conversational agents are currently designed to automate a number of tedious mobile tasks (e.g., to call a taxi), but the possible applications are endless. A domain of particular interest is that of Cultural Heritage, where conversational agents can act as personalized tour guides in, for example, archaeological attractions. The visitors to historical places have a diverse range of information needs. For example, casual visitors have different information needs from those with a deeper interest in an attraction (e.g., - holiday learners versus students). A personalized conversational agent can access a cultural heritage database, and effectively translate data into a natural language form that is adapted to the visitor’s personal needs and interests. The present research aims to investigate the information needs of a specific type of visitors, those for whom retention of cultural content is important (e.g., students of history, cultural experts, history hobbyists, educators, etc.). Embodying a conversational agent enables the agent to use additional modalities to communicate this content (e.g., through facial expressions, deictic gestures, etc.) to the user. Simulating the social norms that guide the real-world human-to-human interaction (e.g., adapting the story based on the reactions of the users), should at least theoretically optimize the cognitive accessibility of the content. Although a number of projects have attempted to build embodied conversational agents (ECAs) for cultural heritage, little is known about their impact on the users’ perceived cognitive accessibility of the cultural heritage content, and the usability of the interfaces they support. In particular, there is a general disagreement on the advantages of multimodal ECAs in terms of users’ task performance and satisfaction over nonanthropomorphised interfaces. Further, little is known about what features influence what aspects of the cognitive accessibility of the content and/or usability of the interface. To address these questions I studied the user experiences with ECA interfaces in six user studies across three countries (Greece, UK and USA). To support these studies, I introduced: a) a conceptual framework based on well-established theoretical models of human cognition, and previous frameworks from the literature. The framework offers a holistic view of the design space of ECA systems b) a research technique for evaluating the cognitive accessibility of ECA-based information presentation systems that combine data from eye tracking and facial expression recognition. In addition, I designed a toolkit, from which I partially developed its natural language processing component, to facilitate rapid development of mobile guide applications using ECAs. Results from these studies provide evidence that an ECA, capable of displaying some of the communication strategies (e.g., non-verbal behaviours to accompany linguistic information etc.) found in the real-world human guidance scenario, is not affecting and effective in enhancing the user’s ability to retain cultural content. The findings from the first two studies, suggest than an ECA has no negative/positive impact on users experiencing content that is similar (but not the same) across different locations (see experiment one, in Chapter 7), and content of variable difficulty (see experiment two, in Chapter 7). However, my results also suggest that improving the degree of content personalization and the quality of the modalities used by the ECA can result in both effective and affecting human-ECA interactions. Effectiveness is the degree to which an ECA facilitates a user in accomplishing the navigation and information tasks. Similarly, affecting is the degree to which the ECA changes the quality of the user’s experience while accomplishing the navigation and information tasks. By adhering to the above rules, I gradually improved my designs and built ECAs that are affecting. In particular, I found that an ECA can affect the quality of the user’s navigation experience (see experiment three in Chapter 7), as well as how a user experiences narrations of cultural value (see experiment five, in Chapter 8). In terms of navigation, I found sound evidence that the strongest impact of the ECAs nonverbal behaviours is on the ability of users to correctly disambiguate the navigation of an ECA instructions provided by a tour guide system. However, my ECAs failed to become effective, and to elicit enhanced navigation or retention performances. Given the positive impact of ECAs on the disambiguation of navigation instructions, the lack of ECA-effectiveness in navigation could be attributed to the simulated mobile conditions. In a real outdoor environment, where users would have to actually walk around the castle, an ECA could have elicited better navigation performance, than a system without it. With regards to retention performance, my results suggest that a designer should not solely consider the impact of an ECA, but also the style and effectiveness of the question-answering (Q&A) with the ECA, and the type of user interacting with the ECA (see experiments four and six, in Chapter 8). I found that that there is a correlation between how many questions participants asked per location for a tour, and the information they retained after the completion of the tour. When participants were requested to ask the systems a specific number of questions per location, they could retain more information than when they were allowed to freely ask questions. However, the constrained style of interaction decreased their overall satisfaction with the systems. Therefore, when enhanced retention performance is needed, a designer should consider strategies that should direct users to ask a specific number of questions per location for a tour. On the other hand, when maintaining the positive levels of user experiences is the desired outcome of an interaction, users should be allowed to freely ask questions. Then, the effectiveness of the Q&A session is of importance to the success/failure of the user’s interaction with the ECA. In a natural-language question-answering system, the system often fails to understand the user’s question and, by default, it asks the user to rephrase again. A problem arises when the system fails to understand a question repeatedly. I found that a repetitive request to rephrase the same question annoys participants and affects their retention performance. Therefore, in order to ensure effective human-ECA Q&A, the repeat messages should be built in a way to allow users to figure out how to ask the system questions to avoid improper responses. Then, I found strong evidence that an ECA may be effective for some type of users, while for some others it may be not. I found that an ECA with an attention-grabbing mechanism (see experiment six, in Chapter 8), had an inverse effect on the retention performance of participants with different gender. In particular, it enhanced the retention performance of the male participants, while it degraded the retention performance of the female participants. Finally, a series of tentative design recommendations for the design of both affecting and effective ECAs in mobile guide applications in derived from the work undertaken. These are aimed at ECA researchers and mobile guide designers.

Item Type: Thesis (PhD)
Research Areas: A. > School of Science and Technology > Computer and Communications Engineering
B. > Theses
Item ID: 12627
Depositing User: Users 3197 not found.
Date Deposited: 08 Nov 2013 10:34
Last Modified: 30 Nov 2022 00:19
URI: https://eprints.mdx.ac.uk/id/eprint/12627

Actions (login required)

View Item View Item


Activity Overview
6 month trend
6 month trend

Additional statistics are available via IRStats2.