Hypertext Information Retrieval 1 Running Head: EXPERT HYPERTEXT INFORMATION RETRIEVAL Expertise and Semantic Grouping in Hypertext Information Retrieval Tasks

The facilitative effect of expertise in hypertext information retrieval tasks (IR) has been widely reported in related literature. However, recent theories of human expertise question the robustness of this result, since previous works have not considered the interaction between user and system characteristics (Vicente and Wang 1998, Vicente 2000). In this study work, the Constraint attunement hypothesis (CAH) is considered in order to predict that effect of expertise in IR would appear only when the user and system characteristics can be combined successfully. Results of 3 experiments revealed that expert users outperform novice users in IR when the elements of a system interface are organized semantically, but not when organized randomly. Results are discussed in the framework of the CAH and the Embodied cognition, supporting the interactive nature of human behaviour in HCI. Hypertext Information Retrieval 3 Expertise and Semantic Grouping in Hypertext Information Retrieval Tasks Hypertext systems consist of a network of linked documents (texts, images, tables...) that users can access using different features like menus or embedded links. Due to its potential in organizing a great amount of information, one of the main tasks that users perform in this system is information retrieval (IR). Researchers have investigated several task and user factors that influence performance in this task (e.g. Chen and Rada 1996). Expertise on the information domain is one of the most important user factors to explain performance in IR: high knowledge users are found to be more efficient and accurate than low knowledge users. This effect has been found repeatedly in hypertexts with different content structure (Shin et al. 1994, McDonald and Stevenson 1998a, Patel et al. 1998), menu organization (Hollands and Merikle 1987), navigation tools (McDonald and Stevenson 1998b) or type of goal (Last et al. 2001). Researchers agree that this effect is due to the fact that experts hold an accurate mental representation of the contents hence allowing them to engage in a more efficient knowledge driven search (e.g. McDonald and Stevenson 1998b). However, the robust effect of this previous knowledge is questioned in recent theories of expertise (Vicente and Wang 1998, Vicente 2000). The authors proposed the Constraint attunement hypothesis (CAH), stating that the effects of expertise on a given task do not always hold, because they depend on the level that the structure of the environment helps the person to use his expertise. In the words of Vicente and Wang (1998: 36): ‘There can be expertise effects when there are goal-relevant constraints (i.e. relationships pertinent to the domain) that experts can exploit to structure the stimuli. The more constrain available, the greater the expertise advantage can be’. Following the CAH, the effects of domain expertise on IR should not always be present, as previous works in the field suggest. They would depend , however, on characteristics of the hypertext system relevant to the task (e.g. menus, interface or structure). Consider the case of a student that has had previous experience with class registration. As far as the registration system follows the task structure the person is used to, he Hypertext Information Retrieval 4 should perform better than a freshman student that has not had previous experience in the domain. But in the case that the system violates some rule of the “registration scheme” (e.g. by using a poor menu organization, like including the “final exams schedule” under the label “grading and enrolment information” instead of “academic calendar”) the effects of expertise on the domain could disappear. Some works assessing the effect of grouping system elements on IR partially support this hypothesis (Halgren and Cooke 1993, Marketta and Saariluoma 2003, Marketta In press). These works have shown that a semantic grouping of the elements of an interface improve performance (e.g. by using the Gestalt laws of similarity of closeness), but also, however, that this improvement disappears when labels are grouped randomly. Materials used in these experiments are relatively simple (e.g. semantic categories such as “domestic animals” or “sports”). Therefore, it can be assumed that most of the participants possessed a high knowledge of the contents before interacting with the system. However, in these experiments the level of expertise of the user was not manipulated, so it can not be affirmed that these effects hold only for expert users, as the CAH would predict. In this present study, we start by considering the CAH expertise theory in order to question the robust effect of previous knowledge in IR. The general hypothesis derived from this framework states that expertise effects would only appear when the external conditions of a system favour their expertise. Therefore, the theory emphasizes both the role of the user and the environment for explaining expertise behaviour. In order to operationalize this hypothesis using concrete predictions on the interaction between the knowledge represented on the user memory (expertise) and that represented on the interface (grouping), we need a theory of mental representation compatible with the CAH view which considers both the user and the environment. The theoretical framework of the “Embodied cognition” is useful for this purpose (Barsalou 1993, Glenberg and Robertson 1999, 2000, Kschak and Glenberg 2000). This theory states that our cognitive system has evolved in order to facilitate our interaction with the environment. For this reason, the world is conceived as a set of possible Hypertext Information Retrieval 5 interactions between the human body and the environment. Therefore, the meaning of an object for a person consists of the things that he/she can do with that object. An object can be used for different purposes, therefore the meaning is flexible and depends on the context (Barsalou 1993). In this sense, it is predicted that knowledge can be successfully activated when properties of the environment (the so called projectable properties) are successfully combined with stored previous experiences related with them (the so called non-projectable properties) (Glenberg and Robertson, 1999, 2000, Kaschak and Glenberg, 2000). Following this theoretical framework, we have derived some hypotheses concerning expert interaction in IR. The main hypothesis states that performance depends on the combination between the non-projectable properties of the user and the projectable properties of the interface. If both are compatible, then interaction would be successful. If they are not compatible, then interaction would be hampered. In addition, two other possibilities can be considered. Firstly, in the case that the interface does not provide projectable properties, the non-projectable properties of the user are expected to guide the interaction. Secondly, in the case that no mental representation about the domain is available, the different projectable properties of the interface would not affect performance, since there is no way to combine them with previous stored experiences with the domain. We tested these hypotheses in three experiments where participants had to search for information in a hypertext system. In experiment 1 and 2, projectable and non-projectable properties of the interaction were assessed in isolation. In experiment 3, both types of properties were assessed conjointly in order to study their interaction. Experiment 1 In experiment 1 we studied the isolated effect of the projectable properties of an interface, varying the types of groupings of labels of the main menu of a hypertext system (semantic grouping, random grouping and no grouping). We hypothesized that semantic grouping would improve Hypertext Information Retrieval 6 performance, whereas random grouping would hamper it. The no grouping condition was used as a control in which no projectable properties were available. Method Participants Forty-two undergraduate students from the University of Granada participated in the experiment for class credits. None of them had had previous experience with the web site used in the experiment. Materials A web site of a scientific meeting was designed for this experiment. The site consisted of 11 pages hierarchically organized. The main page contained links to the rest of the pages. Each link was surrounded by a coloured rectangle (figure 1). Three menus were created, as stated above: a semantically grouped menu, a randomly grouped menu, and an ungrouped menu. For each group of each condition the label colour was identical, and for the ungrouped menu the same colour was used for all labels. Groups from the semantic menu were created according to an expert solution. Four experts in the domain with experience of scientific meetings performed a card sorting task with items from the main menu (Valero and Sanmartin 1999). The two grouping solution was chosen because it had the highest intra-expert agreement. Experts agreed in labelling the two groups as “Bureaucratic” and “Content info”. The web site was written in HTML language, and the navigation behaviour along with the rest of the measures were controlled by a program written PC computers Microsoft Visual Basic 6.0. . --------------------------------FIGURE 1 ABOUT HERE --------------------------------Procedure Participants performed a 20 items search task through the web site. The presentation of the Hypertext Information Retrieval 7 items was divided into two blocks. In each block participants had to search for 10 items which were included in one of the 10 different pages. Therefore, in each block participants had to visit all the pages. Participants had 90 seconds in order to find each item. Design We used a 3 x 2 Mixed Factorial design. Menu (semantic, random and ungrouped) was the between-participants variable and Block (1 and 2) was the within-participant variable. Performance carried out in the search task was measured by two dependent variables: response time and lostness. Lostness was defined following the formula proposed by Smith (1996): L = (N/S -1) 2 + (R/N -1) 2, were N is the number of different nodes visited while performing the search task; S is the total number of nodes visited while performing the search task; and R is the number of nodes needed in order to accomplish the search task. Lostness index ranges between 0 and 1. The greater the value, the greater the lostness. Results Significance level for all experiments was set at alfa 0.05. Different performance data was submitted to a 3 x 2 ANOVA with Menu (semantic, random and ungrouped) as the between-participants variable and Blocks (1 and 2) as the within-participant variable. Results are summarized in table 1. Response time. Response times are reported in seconds. Results showed a main effect of Menu, F(2, 39) = 12.95, Mse = 66.22. Follow up comparisons revealed that participants of the semantic menu performed faster (M = 15.83) than those of the ungrouped (M = 22.38) and random menus (M = 26.57). Differences between the random and ungrouped conditions were close to significance, p = 0.07. The main effect of Blocks was also significant, F(1, 39) = 91.73, Mse = 32. Participants performed slower in Block 1 (M = 27.53) than in Block 2 (M = 15.66). In addition, the interaction between Menu and Stage was also significant, F(2, 39) = 7.35, Mse = 32. Follow up analysis revealed that in Block 1, Hypertext Information Retrieval 8 participants of the semantic menu were faster (M = 19.34) than those of the ungrouped (M = 27.54), and both of them were faster than those of the random (M = 35.72). In Block 2, participants of the semantic menu were faster (M = 12.32) than those of both the ungrouped (M = 17.23) and random conditions (M = 17.42). Lostness. Data is reported with the L value of lostness (Smith 1996). Results showed a main effect of Menu, F(2, 39)= 10.99, Mse = 0.03. Follow up comparisons revealed that participants on the semantic menu got lost fewer times (M = 0.21) than those using the ungrouped menu (M = 0.35). Moreover, both of them got lost fewer times than those using the random menu (M = 0.4). The main effect of Block was also significant, F(1, 39) = 6.27, Mse = 0.02. Participants got lost more times in Block 1 (M = 0.36) than in Block 2 (M = 0.28). Follow up comparisons revealed that this effect was only observed in the participants of the random menu. The first order interaction between Main page and Block was not significant, p < 0.35. However, it is interesting to note that participants of the three menus only differed in Block 1 (semantic M = 0.24; ungrouped M = 0.36; and random M = 0.47), F (1, 39) = 6.98, Mse = 0.02 for the comparison between the semantic and ungrouped conditions, F (1, 39) = 4.44, Mse = 0.02 for the comparison between the ungrouped and random conditions; however in Block 2 only the participants of the semantic menu got lost fewer times than the other two conditions (semantic M = 0.19; ungrouped M = 0.33; and random M = 0.33), F (1, 39) = 8.04, Mse = 0.03. --------------------------------TABLE 1 ABOUT HERE --------------------------------Discussion Results of experiment 1 replicated previous results concerning the facilitative effect of semantic grouping and the inhibitory effect of random grouping in information retrieval tasks (Marketta and Saariluoma 2003, Marketta In press). In addition, performance results show the distinct effect of Hypertext Information Retrieval 9 varying the projectable properties of the menu. Participants of the semantic condition perform faster and got lost fewer times than those of the ungrouped condition, moreover they perform better than those of the random condition. The facilitation and inhibition of the semantic and random conditions can be explained due to the combination of the projectable properties of the menu and the mental representation of the participants. In the semantic condition, both properties are compatible, whereas in the random condition, projectable properties can not be combined with the mental representation of the user. In the ungrouped conditioned menu, no projectable properties concerning the organization of the menu are available, so the participant has to activate its knowledge about the system without any clue from the interface. In this case, there is no such facilitation like that observed in the semantic condition, nor an interference like that observed in the random condition. However, in experiment 1 the level of subjects expertise between conditions was not controlled, so we were not able to assume that these results applied to participants with different levels of expertise. For this reason, in the following experiments we explored the effect of this variable; both isolated (experiment 2) and combined with different types of grouping (experiment 3). Experiment 2 In experiment 2 we studied the isolated effect of the non-projectable properties of IR, selecting participants with different levels of domain knowledge (novices, intermediates and experts), who in turn had to interact with an ungrouped menu. We hypothesized that domain experts would perform better than novice users. In addition, we predicted that intermediate users would behave in a more similar fashion to novice users due to the lack of projectable properties of the interface that could facilitate the activation of their knowledge. Method Participants Eighteen students from the University of Granada participated in the experiment. Twelve were Hypertext Information Retrieval 10 undergraduates participating for class credits. Six were graduate students participating voluntarily. None of them had had previous experience with the hypertext used in the experiment. Materials and procedure The menu used in experiment 2 consisted of the same ungrouped menu used in experiment 1 (figure 1c). The procedure was similar to that used in experiment 1 except for two changes in the search task. In experiment 2, the trial time limit was established at 60 seconds (instead of the 90 seconds used in the first experiment). We predicted that with less time participants would not be able to find all targets, therefore we would be able to use the variable number of targets found as a complementary variable. The second change consisted of an increase in the number of trials from 20 to 30 trials. Design We used a 3 x 3 Mixed Factorial design. Expertise (novice, intermediate and expert users) was the between-participants variable and Block (3 levels) was the within-participant variable. Expertise of participants was established by considering the number of times they had attended a scientific meeting: novice users 0 times, intermediate users 1 to 4 times, and expert users more than 4 times. Performance in the search task was measured using three dependent variables: the number of targets found, the response time and lostness. Results Different performance data was submitted to a 3 x 2 ANOVA with Expertise (novice, intermediate and expert) as the between-participants variable and Blocks (3 levels) as the withinparticipant variable. Analysis of the response time and lostness was only conducted on the trials where the participant had found the target. Results are summarized in table 2. Number of targets found. The only reliable difference for this variable was the main effect of Blocks, F (2, 30) = 10.30, Mse = 0.71. Participants found fewer targets in block 1 (M = 8.4), than in 2 Hypertext Information Retrieval 11 (M = 9) and fewer than in 3 (M = 9.7). Response time. Results showed a close to significant main effect of Expertise, F (2, 15) = 3.15, Mse = 33.23, p = 0.07. Follow up comparisons made in order to test our hypotheses revealed that novice users (M = 14.7) and intermediate users (M = 13.5) together performed slower than experts (M = 10), F (1, 15) = 5.95, Mse = 33.23. In addition, the main effect of Blocks was significant, F (2, 30) = 14.37, Mse = 15.08. Participants were slower in the first block (M = 16.7), than in the second (M = 11.36), and third (M = 10.21). The interaction between expertise and blocks was not significant. Lostness. No significant differences were found for this variable. --------------------------------TABLE 2 ABOUT HERE --------------------------------Discussion Results of experiment 2 support partially the hypothesis that in the absence of projectable properties in the interface, experts perform better in an information retrieval task than novice and intermediate users. However, this hypothesis was only confirmed by the response time, and not by the number of targets found or by the lostness value. A possible explanation for these results is the fact that no projectable properties are available in the task, therefore the expert participants are not able to take advantage of their knowledge. In a task in which experts can successfully combine their knowledge with the projectable properties of the interface, these differences should be greater for all the performance variables. Following this line of thought, in experiment 3 we assessed the combined effect of expertise level and the type of grouping. Experiment 3 In experiment 3 we studied the combined effect of both the projectable and non-projectable properties of an information retrieval task. We hypothesized that the facilitative effect of semantic Hypertext Information Retrieval 12 grouping would only be effective for knowledgeable participants (expert and intermediate users), and therefore they would perform better than novice users. In addition, we hypothesized that the inhibitive effect of random grouping would only affect knowledgeable participants (expert and intermediate users), and therefore the differences expected in the semantic condition would disappear. Method Participants Thirty-six students from the University of Granada participated in the experiment. Twenty-two were undergraduates participating for class credits. Fourteen were graduate students participating voluntarily. None of them had had previous experience with the hypertext used in the experiment. Materials and procedure The menus used in experiment 3 consisted of the semantic and random menu used in experiment 1 (figure 1a and 1b). The procedure was identical to that used in experiment 2. Design We used a 3 x 2 x 3 Mixed Factorial design. Expertise (novice, intermediate and expert users; defined as in experiment 2) and Menu (semantic and random) were the between-participants variables and Blocks (3 levels) was the within-participant variable. Performance on the search task was measured by three dependent variables: the number of targets found, the response time and lostness. Results Different performance data was submitted to a 3 x 2 x 3 ANOVA with Expertise (novice, intermediate and expert) and Menu (semantic and random) as between-participants variables and Blocks (3 levels) as within-participant variable. Analysis of the response time and lostness were conducted only on the trials where the participant had found the target. Due to the complexity of the design, we have focused on the comparisons conducted in order to test our hypothesis. Results are summarized in table 3. Hypertext Information Retrieval 13 Number of targets found. Results revealed a main effect of Blocks, F (2, 60) = 15.65, Mse = 0.7. Participants performed worse in the first block (M = 8.6) than in the second (M = 9.5) and third (M = 9.7). In addition, supporting our hypotheses results showed that in the semantic menu novice users found less targets (M = 26,6) than intermediate users (M = 30) and experts (M = 29), F (1, 30) = 8.6, Mse = 1.41; but this difference disappeared in the random menu (novice users M = 27.2; intermediate users M = 26.5; experts M = 28.3), F < 1. Response time. Results showed a main effect of Blocks, F (2, 60) = 9.72, Mse = 12.81. Participants performed slower in the first block (M = 13.4), than in the second (M = 10.3) and third (M = 9.8). In addition, supporting our hypotheses results revealed that in using the semantic menu novice users were slower (M = 13.4) than intermediate users (M = 9.4) and experts (M = 8), F (1, 30) = 7.02, Mse = 39.68; but that there were no differences observed in the random menu, (novices M = 12.3; intermediates M = 12; experts M = 11.9), F < 1 (see figure 2). --------------------------------FIGURE 2 ABOUT HERE --------------------------------Lostness. Results showed no practice effect for this variable, F < 1. Analysis supporting our hypothesis was close to significance, F (1, 30) = 3.26, Mse = 0.04, p = 0.08. In the semantic condition, novice users got lost more (M = 0.23) than intermediate users (M = 0.16) and experts (M = 0.1); whereas these differences were not found in the random condition (novice users M = 0.18; intermediate users M = 0.18; expert users M = 0.15), F < 1. --------------------------------TABLE 3 ABOUT HERE --------------------------------Discussion Hypertext Information Retrieval 14 Results from experiment 3 support the hypothesis that the effect or domain expertise on IR depends on the combination of user knowledge and system characteristics. When the latter are compatible with expert mental representation, knowledgeable users find more information, in less time and in a more efficient way than novice users. The expertise effect disappears when the characteristics of the system are not compatible with the knowledge of the expert user. In addition, for participants without previous knowledge in the domain, there are no effects of the groupings, even after practicing with the system. These results can be explained using the framework of the CAH and the Embodied cognition. In the case of knowledgeable participants, their non-projectable properties can be combined successfully with projectables properties of a semantic grouping interface, thus improving performance. However, when interacting with a randomly grouped menu, the projectable properties of the interface will interfere with their non-projectable properties, disabling a successful combination, therefore hampering performance. In the case of novice participants, the projectable properties of the semantic and random menu can not be combined with non-projectable properties in the domain, and therefore they do not perform any different effect on the participant's performance. General discussion Previous works in the related literature have extensively reported the positive effect of expertise in IR (Hollands and Merikle 1987, Shin et al. 1994, McDonald and Stevenson 1998a, 1998b, Patel et al. 1998, Last et al. 2001). This result has been explained by the fact that experts have an accurate mental representation of the contents that guide their search through the system. However, recent theories of expertise question the generalization of this explanation (Vicente and Wang 1997, Vicente 2000). Expertise effects are expected to appear only when the expert mental representation can be matched with the information presented on the interface (e.g. the way the items on a menu are organized). Results described in this study support this idea, showing that knowledgeable users Hypertext Information Retrieval 15 outperform novice users in IR when the elements of an interface are semantically organized, however not when they are randomly organized. This study has benefited from the application of psychological theories emphasizing the interactive nature of human cognition (CAH and Embodied cognition). User interaction with a system can not be explained by only focusing on his / her characteristics, nor only on system ones characteristics. On the contrary, this behaviour must be explained by the interaction of both human and system characteristics (Cañas et al. In press). Previous results reporting the isolated effects of expertise and semantic grouping in IR has been widely reported on the literature (see introduction) and replicated in experiments 1 and 2. However, the complete picture of expert behaviour in IR is not explained until both human and system characteristics are explored in conjunction. Results reported in ths study have relevant implications for the design of hypertext systems. A well established design guideline asserts that system interfaces must map the way the user organizes its knowledge of the system in memory (his / her mental model of the system) (Norman 1988). Our results support this guideline in hypertext IR, but only if the system is devoted to domain expert users. If the system is mainly dedicated to novice domain users, the effort of engaging an analysis of the user’s mental model can be avoided. In the experiments reported here, only the intermediate and expert users have benefited from mental model mapping, but it has had no influence on novice users. As exemplified in the above experiments, a simple way of mapping the user’s mental model onto a hypertext could be to use the Gestalt laws for facilitating perceptual groupings of related elements (e.g. Marketta and Saariluoma 2003). The grouping of related elements can be done by using the same colour for each related group, by arranging related elements close to each other, or by connecting related elements by a graphical element (lines, boxes...). However, in the case that the mapping of the user’s mental model is not done, attention must be paid to the use of design elements in the interface which can possibly induce an arbitrary grouping of elements. In this case, the Hypertext Information Retrieval 16 performance of intermediate and expert domain users in IR can be hampered. ReferencesBARSALOU, L.W., 1993, Flexibility, structure, and linguistic vagary in concepts: manifestations of acompositional system of perceptual symbols. In Theories of memories, edited by A.C. Collins, S.E.Gathercole, and M.A. Conway. (London: Erlbaum), pp. 29-101.CAÑAS, J.J., SALMERÓN, L., and FAJARDO, I., In press, Toward the analysis of the interaction in the jointcognitive system. In Future interaction design, edited by A. Pirhonen, H. Isomäki, C. Roast and P.Saariluoma (London: Springer-Verlag).CHEN, C., and RADA, R., 1996, Interacting with hypertext: a meta-analysis of experimental studies.Human-Computer Interaction, 11, 125-156.GLENBERG, A. M. and ROBERTSON, D. A., 1999, Indexical understanding of instructions. DiscourseProcesses, 28, 1-26.GLENBERG, A. M., and ROBERTSON, D. A., 2000, Symbol grounding and meaning: a comparison ofhigh-dimensional and embodied theories of meaning. Journal of Memory & Language, 43, 379-401.HALGREN, S. L., and COOKE, N. J., 1993, Towards ecological validity in menu research. InternationalJournal of Man-Machine Studies, 39, 51-70.HOLLANDS, J. G., and MERIKLE, P. M., 1987, Menu organization and user expertise in informationsearch tasks. Human Factors, 29, 577-586.KASCHAK, M. P., and GLENBERG, A. M., 2000, Constructing meaning: the role of affordances and grammatical constructions in sentence comprehension. Journal of Memory & Language, 43, 508-529.LAST, D., O'DONNELL, A., and KELLY, A., 2001, The effects of prior knowledge and goal strength onthe use of hypertext. Journal of Educational Multimedia & Hypermedia, 10, 3-25.MARKETTA, N., In press, Layout arrangement and pop-up labels: effects on search. ScandinavianJournal of Psychology. Hypertext Information Retrieval 17 MARKETTA, N., and SAARILUOMA, P., 2003, Layout attributes and recall. Behaviour & InformationTechnology, 22, 353-363.MCDONALD, S., and STEVENSON, R. J., 1998a, Effects of text structure and prior knowledge of thelearner on navigation in hypertext. Human Factors, 40, 18-27.MCDONALD, S., and STEVENSON, R. J., 1998b, Navigation in hyperspace: An evaluation of the effectsof navigational tools and subject matter expertise on browsing and information retrieval in hypertext.Interacting with Computers, 10, 129-142.NORMAN, D. A., 1988, The psychology of everyday things (New York: Basic Books).PATEL, S. C., DRURY, C. G., and SHALIN, V. I., 1998, Efectiveness of expert semantic knowledge as anavigational aid within hypertext. Behaviour & Information Technology, 17, 313-324.SHIN, E. C., SCHALLERT, D. L., and SAVENYE, W. C., 1994, Effects of learner control, advisement, andprior knowledge on young students’ learning in a hypertext environment. Educational TechnologyResearch and Development, 42, 33-46. SMITH, P.A., 1996, Towards a practical measure of hypertext usability. Interacting with Computers, 4,365-381. VALERO, P., and SANMARTIN, J., 1999, Methods for defining user groups and user-adjusted informationstructures. Behaviour & Information Technology, 18, 245-259. VICENTE, K. J., 2000, Revisiting the constraint attunement hypothesis: reply to Ericsson, Patel, andKintsch (2000) and Simon and Gobet (2000). Psychological Review, 107, 601-608.VICENTE, K. J., and WANG, J. H., 1998, An ecological theory of expertise effects in memory recall. Psychological Review, 105, 33-57. Hypertext Information Retrieval 18 Figure 1. Different menu interfaces used in experiments 1-3 (the originals were constructed inSpanish). (A) Semantic menu; (B) Random menu; (C) Ungrouped menu. Hypertext Information Retrieval 19 Expertise:NovicesExpertise:IntermediatesExpertise:ExpertsMenu: SemanticResponsetime 678910111213141516 Block: 123Menu: RandomBlock: 123 Figure 2. Response times in experiment 3 by expertise (novices, intermediates and experts) and typesof menu (semantic and random). Hypertext Information Retrieval 20 Table 1. Means and standard deviations (in parenthesis) for performance variables of experiment 1.---------------------------------------------------------------------------------Type of menuSemantic Random UngroupedDependentVariable---------------------------------------------------------------------------------Response time 15,8 (6,1) 26,5 (7) 22,3 (7,8)Lostness 0,21 (0,12) 0,4 (0,16) 0,35 (0,14) Table 2. Means and standard deviations (in parenthesis) for performance variables of experiment 2.---------------------------------------------------------------------------------ExpertiseNovice Intermediate ExpertDependentVariable---------------------------------------------------------------------------------Targets found 25,6 (4,8) 27,1 (3,4) 28,5 (1,7)Response time 14,7 (5) 13,5 (4,2) 10 (2,6)Lostness 0,26 (0,18) 0,16 (0,13) 0,17 (0,17) Table 3. Means and standard deviations (in parenthesis) for performance variables of experiment 3.----------------------------------------------------------------------------------------------------------------------------------------Expertise & Type of menuNoviceIntermediateExpertSemantic Random Semantic Random Semantic RandomDependentVariable----------------------------------------------------------------------------------------------------------------------------------------Targets found 26,6 (3,7) 27,2 (2,9) 30 (0)26,5 (2,9) 29 (1,6) 28,3 (1,5)Response time 13,3 (6,1) 12,2 (6,2) 9,4 (3)12 (3)8 (2,8)11,9 (3,6)Lostenss 0,23 (0,2) 0,18 (0,14) 0,16 (0,15) 0,18 (0,13) 0,1 (0,11) 0,15 (0,17) Hypertext Information Retrieval 21