Navigating in a Virtual Environment with Model-Generated Support Herre van Oostendorp
Saraschandra Karanam
Dept. of Information and Computing Sciences Utrecht University Princetonplein 5 3584CC Utrecht The Netherlands
Xerox Research Center India Neil Rao Towers 118, Road #3, EPIP Phase-I, Whitefield, Bangalore 560 066, India
[email protected]
[email protected]
ABSTRACT Though the cognitive processes controlling user navigation in virtual environments as well as in websites are similar, cognitive models of web-navigation have never been used for generating support in virtual environment navigation. We created a simulated 3D building of a hospital and presented users various navigation tasks under two conditions: a control condition and a modelgenerated support condition. Mean task-completion time and disorientation were recorded. It was found that the cognitive model used can simulate the navigation behavior of participants and also that with model-generated support participants took significantly less time to reach their destination and were significantly less disoriented. The impact of providing model-generated support on disorientation was especially higher for users with low spatial ability. We demonstrated that it is possible to generate tools for navigation in virtual environments using cognitive models developed for web-navigation. Author Keywords Web-navigation; automated tools; virtual navigation; cognitive model; virtual environment. ACM Classification Keywords H.5.4 [ Hypertext/Hypermedia] Navigation General Terms Navigation
INTRODUCTION Virtual Environment (VE) is a term applied to computer simulated environments of real and imaginary worlds. There has been an increasing use of this technology in many fields like archaeology to reconstruct old heritage sites, caves, monuments and sculptures, health care to simulating complex surgeries, and therapeutics to treat phobias and certain motor and limbic impairments. VEs or serious games are also successfully used in the domain of training pilots for emergency situations like crash landing, or triage procedures for medical first responders etc. VE is also highly popular among movie and game enthusiasts. However, due to information overload it is sometimes difficult for users to find their way and to know how to proceed. Particularly novice users or users with limited cognitive abilities may experience problems in navigating to their target destination [1, 2, 3, 4]. With so much popularity and usage, it Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from
[email protected]. ECCE '13, August 26 - 28 2013, Toulouse, France Copyright 2013 ACM 978-1-4503-2251-5/13/08…$15.00. http://dx.doi.org/10.1145/2501907.2501948
is important for VE developers to reduce the user’s cognitive load and maximize the effectiveness, efficiency and user satisfaction. In this paper, we focus on the aspect of navigation in virtual environments. First, we will argue that problems in navigating in VEs are similar to problems web users experience and discuss some tools based on cognitive models that are developed for web-navigation. Finally we will demonstrate in the following section via a behavioral study the usefulness of a cognitive model of web-navigation for navigation in a 3D virtual environment.
WEB-NAVIGATION: COGNITIVE MODELS AND SUPPORT TOOLS The complexity and non-linear nature of documents on the Web is known to cause a lot of problems for web users. Users often report experiencing a feeling of being lost or not knowing where they are [5]. Users are known to employ strategies when they navigate on the web, surfing for content. They follow what is called information scent of proximal cues (hyperlinks, images and other content) that direct them to the distal target page. Information scent is defined as the imperfect perception of the cost or gain associated with taking a particular action such as clicking on a particular hyperlink on a website [6]. Users try to maximize their gain by following only those proximal cues that give highest information scent. Inspired by this concept, several cognitive models of web-navigation have come up in recent years. We describe briefly one of such models called Comprehension-based Linked Model of Deliberate Search or CoLiDeS [7, 8, 9]. According to CoLiDeS, navigating on a website involves four processes: parsing the webpage into 5-10 high level schematic regions, focusing on one of these schematic regions, comprehension of screen objects within that schematic region, and finally selecting the most appropriate actual object in that schematic region. CoLiDeS selects the most appropriate object by measuring the semantic relatedness of various screen objects in the selected schematic region with the user goal. This measure provides the operationalization of information scent. Latent Semantic Analysis technique is used to compute the semantic similarity [10]. Users tend to follow the objects with the highest information scent and click on the corresponding link label. This process is repeated for every new screen page, until the user reaches the target destination. Several studies have shown the validity of this modeling approach and have also supported usefulness of the LSA technique [9, 10, 11, 12, 13, 14]. Similarities in navigating the Web and virtual environments We want to argue that there are some interesting common processes and problems in navigating in VEs and web search. Focusing of attention, planning, and knowledge activation from
memory and keeping it in working memory are cognitive processes that are crucial in both activities [15]. Also there are numerous instances in 3D environments, just as in browsing the Web (in fact also in real life), when we are overwhelmed with information overload. Yet, with some effort and focusing of attention, we are often able to come out of the cognitive overload caused by it and reach our target. Examples of such VE scenarios could be navigating in a complex building like a virtual library and looking for a section on “garden flowers” (first looking for a section within nonfiction literature, next natural science, and then biology, etc), or navigating in a virtual airport looking for the shortest route from one gate to another (is it an internal, national flight, or international flight, do I have to pick up first my luggage, when to look for customs, etc). In all of these scenarios, the most common approach is to look for some information boards that guide us towards our target destination. It is interesting to note that the parsing, focusing, planning and knowledge activation strategies we employ in such situations are very similar to the strategies employed by a user navigating on the internet. We weigh various options, develop a vague estimate of the cost / gain associated with each choice and pick the one that best suits us according to the context and our goal, as put forth by Pirolli and Card in their Information Foraging Theory [6]. Let us take the example of a patient navigating in a 3D model of a hospital building to elaborate on this point: Suppose a user, e.g. a patient, is standing in the model at the entrance of the hospital and wants to find the office of a doctor at the neuropsychology department. Around him, he can see a number of people (patients, doctors, nurses), objects like a reception desk, rooms, chairs, stairs, etc. Many information boards with directions for various sections of the hospital are visible. This is equivalent to parsing phase of CoLiDeS. He first has to decipher all this information and choose to focus on one region that suits him the best, maybe an information board. This phase is equivalent to the focusing phase of CoLiDeS. As he approaches the information board in the model and moves the mouse in that direction, the labels on it become more and more clear. He reads them and understands them. This is equivalent to comprehension and elaboration phase of CoLiDeS. Finally, the patient picks one label that is most relevant to his/her goal. This phase is equivalent to selection phase of CoLiDeS. As he crosses that information board and moves deeper into the building, he is faced with similar scenarios again, that is, more furniture, people and more information boards. The above explained processes are then repeated by the user to move further inside the building. Thus, the four phases of CoLiDeS are repeated in cycles until he reaches the target destination, e.g. encountering the target label Depression Center. Often we don't know exactly in the beginning where we have to be. We need context to decide that. That is also the reason that web browsing can be very useful compared to information search by means of search engines which directly bring you to the requested information, at least if you precisely know the goal [16]. Also in navigating a virtual environment it frequently occurs that the goal is ill- defined, and that exploration of the environment has to make clear where the environment can satisfy the goal. This process is called by Pirolli and Card information foraging [6]. Finally, if the information boards and entries (or labels) on the boards are well designed, as a user moves closer to the destination, the labels will get more and more specific. This is also observed in a website: as a user navigates closer to the target page, the hyperlinks get more and more specific.
Tools for browsing and searching the Web For browsing and searching on the Web, several tools have been developed based on cognitive models of web- navigation [11, 12]. ScentTrails [16, 17, 18] for instance brings together the strengths of both browsing and searching behaviours. It operates as a proxy between the user and the web server. In ScentTrails, the user enters keywords representing the portion of his or her information goal that is initially known and amenable to search. ScentTrails then annotates the hyperlinks of web-pages with search cues: an indication that a link leads to content that matches the search query. This annotation is done by visually highlighting links to complement the browsing cues already embedded in each page. The degree to which links are highlighted is determined by an information scent algorithm. It has been found that with ScentTrails running, users could finish their tasks quicker than in a normal scenario. Bloodhound [19] predicts how typical users would navigate through a website hierarchy given their goals. It combines both information retrieval and spreading activation techniques to arrive at the probabilities associated with each hyperlink that specify the proportion of users that will navigate through it [20, 21]. Bloodhound takes a starting page, few keywords that describe the user- goal, and a destination page as input. It then automatically infers the usability of a website. It also provides usability metrics that tell how easy it is to accomplish the given user- goal. While all the above tools either predict user behavior or help in evaluating a website, some of these models have also been used to provide navigation support in the form of auditory cues or highlighted hyperlinks. Van Oostendorp and Juvina [13] and Van Oostendorp, Karanam and Indurkhya [22] performed simulations of user navigation behavior with extended versions of the CoLiDeS model [see also 23]. The navigation support they offered was based on simulations of successful paths, that is, the links chosen by the model and leading to the requested information were subsequently emphasized to the user. They found that the number of clicks was significantly less, users navigated in a more structured manner and task performance was higher with modelgenerated support, and this was found more prominent in users with low spatial abilities. In several of their studies spatial ability appeared to correlate substantially (at least .49) with task performance like answering search questions. See also [24, 25]. They interpret these findings as indicating common cognitive processes between spatial ability and web-navigation like being able to construct an information space and operate on it, such as performing transformations or rotations in that space.
BEHAVIORAL STUDY NAVIGATING IN A VIRTUAL ENVIRONMENT None of the tools mentioned has attempted to explore the application of cognitive models of web-navigation in virtual navigation, as far as we know. We want to explore in this study whether we can apply successfully the CoLiDeS modeling to a VE, and also test whether we can successfully provide support that is generated by the model. We have already described the similarities in cognitive processes in navigation on the Web and navigation in virtual environments. Due to these similarities in cognitive processes, we hypothesize that the cognitive models of web-navigation can be used to generate support for navigation in virtual environment as well. The models can predict at each step, the correct hyperlinks (or labels in the case of virtual environments) for a given goal. We think that improvement in navigation performance of users because of highlighting the model-predicted labels in a virtual environment is an indication of
the soundness of the model. We explore this by designing a 3D model of a hospital building and giving navigation tasks to users under two conditions: a control condition where users do not receive any support and a support condition in which users receive model-generated support in the form of highlighted labels. The labels to highlight are algorithmically computed based on the CoLiDeS model. We assume that for a given goal, at each decision making point, each label is equivalent to a hyperlink in a web site. Thus, we assume that a user would choose that label which is closest and is most relevant to his/her goal. The model estimates relevancy by computing semantic similarity between the user goal and each of the labels using Latent Semantic Analysis (LSA) [10, 12]. The label with the highest semantic similarity measure is supposed to be most relevant to the goal and is highlighted. This process is repeated for every decision point on the path, given a certain goal. It corresponds to the most central algorithm of the CoLiDeS model. We measure the task completion times and the path taken by the user to reach the target location. The main research questions involve, thus, firstly whether we can simulate or predict navigation behavior of participants in a VE and, secondly, whether we can find a positive influence of model-generated support on performance (time and disorientation). As in [24 and 25], we also examine the role of spatial ability and check whether also here support is particularly helpful for low spatial ability participants. Finally, we distinguish different levels of complexity of navigation tasks and study the impact of support in performing high complex tasks.
Figure 1: 3D environment showing some information boards.
METHOD Participants Twenty-four graduate and under-graduate students from Utrecht University, Utrecht, and International Institute of Information Technology at Hyderabad, were randomly assigned to the two conditions, each condition 12 participants. The mean age of the participants was 24 (SD=3.2). Based on the median performance on a spatial ability test both groups were divided into subgroups of 6 participants (below and above median spatial ability score). Material and Apparatus Virtual Environment: We used the source code of the game “Half Life 2” and adapted the environment so that it looked like a hospital. Figure 1 shows a snapshot of the environment. University Medical Centre Utrecht was used as source of inspiration. Based on many pictures of the interior, particularly of the information boards, we constructed a layout of the building and its sections. There were three floors and four different sections in each floor. First floor had Skin, Bone and Genetics, second floor had Surgery, Cancer, Brain and Laboratories and third floor had Allergy, Fertility Care and Infectious Diseases.
Figure 2a: Example information board, control condition. We assume that patients navigating to different sections of the hospital building use these labels. Most labels also come with arrows showing the direction in which the section is. The support condition labels - in red color with a red arrow- were created by computing the semantic similarity between the goal and the labels using LSA (Latent Semantic Analysis) developed by [10] and highlighting the label with the highest cosine value. We used the semantic space ‘tasaALL’ provided by LSA (http://lsa.colorado.edu). This space is meant to represent the knowledge and vocabulary levels of first year university students. Figures 2a and 2b show an information board for both conditions in the VE. In figure 3 we zoomed in on the labels for both conditions for a specific goal. For instance, because the label Surgery, Cancer, Brain, Laboratories appeared to have the highest semantic similarity (.43) with this goal, it was highlighted in red with a red arrow. This process was repeated for all labels lying on the path to the goal. It is important to note that the labels that were emphasized are specific for the respective goals and are generated by the cognitive model. The sequence of labels that were (correctly) chosen for the goal in this concrete example, were:
Surgery, Cancer, Brain, Laboratories, then Brain, then Patient Services, and finally Depression Center.
Table 1: Instructions for the navigation tasks You are at the hospital main entrance. Imagine you have fractured your ankle and foot bones while playing football. Find your way to a doctor who can treat you. You are a PhD student working in the ‘clinical genetics’ section of the department of genetics. Find your way to your section. Imagine that you self-loath, you cannot concentrate; you are irritated and short tempered. You suffer from acute depression. Your doctor referred you to a psychologist in the Brain Center. Visit him. You are at the hospital main entrance. Your father is admitted in the ICU after undergoing a heart surgery. You want to see your cardiologist and take post- operative advice.
Figure 2b: Example information board, in support condition.
You are suffering from liver cancer and you should undergo a chemotherapy session. Go to the liver cancer section of the hospital. You are at the hospital main entrance. Find your way to the chemistry lab of laboratories section. Your kid has been developing frequent and severe infections. Your doctor suspects some kind of viral infections. Go to the Viral Diseases section.
Figure 3: Labels in the control and support conditions for the above goal. Navigation Tasks: A total of nine navigation tasks spread across different sections of the hospital were constructed. Table 1 lists all navigation tasks.
You suddenly developed asthmatic symptoms after eating food. Your doctor advises you to go to the Food Allergy section of Allergy department and take a chest specialist advice. Go to the section. Your wife suffers from infertility. You would like to visit the donor section of fertility department to see if there are mothers who can donate their egg. Visit the section.
An example task for a section on infectious diseases could be “Your kid has been developing frequent and unusually severe infections. Your family advisor suspects some kind of immune deficiency. Go to the “Primary” immunodeficiency section and visit a doctor there. We defined navigation tasks differing in task complexity. For each goal, the number of labels at which a user should take a decision was computed. Using this number, three levels of task complexity were defined: Level 1: 0 < number of labels < 6 Level 2: 6 ≤ number of labels ≤ 8 Level 3: 9 ≤ number of labels ≤ 11 Measures The dependent variables were mean task-completion time and mean disorientation. Task-Completion Time: The time taken by the user to reach the target location was measured, starting from pressing the OK button after having read and understood the task instruction. There was no time limit. Disorientation: An objective measure of lostness was computed
using Smith's (1996) L measure [26]. L = √((N/S – 1)² + (R/N – 1)²) Where: R = number of nodes required to finish the task successfully (thus, the number of nodes on the optimal path); S = total number of nodes visited while searching; N = number of different nodes visited while searching. While a node is a page in a web navigation scenario, in our environment, we defined various intermediate points on the correct path to the target location as node. We noted whether participants passed these points (the labels). The same formula was applied here.
mean task-completion times and disorientation scores were then analyzed separately for both spatial ability groups for each complexity level, in both versions of the 3D model, in a 3x2x2 mixed ANOVA design. Task-Completion Time: The main effect of condition (version) was significant F(1,20) = 15.12, p
