440 likes | 452 Views
This study evaluates the effectiveness of query-biased summaries in web search. Participants use different search interfaces and complete tasks to measure satisfaction and completion time. The findings suggest that query-biased summaries improve search effectiveness.
E N D
i213: User Interface Design & Development Marti Hearst March 5, 2007
Example Study • Sample study by White et al. 2002 • Studying query-biased search results summaries • First did an informal assessment to determine responses to the state of the art • 6 participants • Compared AltaVista & Google • Versions from 2001 or maybe 2000 • Google’s summaries were query-biased, AltaVista’s weren’t • Ranking wasn’t as good then • Findings: • Summaries were ambiguous and too short • First thing they saw was the hit count – discouraging • Had to scroll to see more than a few results • Main conclusion: the document summaries were not descriptive enough
Study Goals • Evaluate a new form of query-biased summaries for web search • Hypothesis: • The presence of query-biased summaries will improve search effectiveness
Experiment Design • Independent Variable: • Search Interface • Levels: • Google, Google + summaries, AV, AV + summaries • Task types • Levels: 4 different tasks • Dependent Variables: • Participant satisfaction • Task completion success • Task completion time
Blocking • Number of participants: 24 • Within-participants • They each use all 4 interfaces • They each do 4 tasks • They control for: • Effects of task • (some harder than others) • Effects of order of exposure to system • (seeing one can influence the effects of seeing the next) • They do not control for: • Order of task
Latin-Square Design • Start with an ordering • Rotate the order, moving one position per line
Latin Square Design T1 T2 T3 T4 G G+ A A+ # partic. Per Cond. 6 G+ A A+ G 6 A A+ G G+ 6 A+ G G+ A within partic. design 6
Latin Square Design • Start with an ordering • Rotate the order, moving one position per line • Note that this doesn’t give you every possible ordering! • (e.g., don’t see AV right after G) • The hope is the outcome isn’t that sensitive to ordering
Study Procedure • Participants came by, one at a time. • Session lasted about 1.5 hours • Procedure: • Introductory orientation session • Background questionnaire • The following 4 steps for each task: • Short training session with the new system • Receive a hard-copy task description • 10 minute search session • Post-search questionnaire • Final questionnaire • Information discussion (optional)
Study Procedure • Data collection • Questionnaires • 5 point Likert scales • 3 on task, 4 on search process, 4 on summaries • Think-aloud • Automatic logging • # docs returned • # summaries requested and returned • # results pages viewed • Time for each session
Subjective Results • No interaction effects for task • All groups preferred the enhanced summaries • All groups feel they benefited from the summaries • G + enhanced summaries was significantly different from the rest on 3 out of 4 (relaxing, interesting, restful) • Except for easy/difficult, where there was no difference • System ranking in the final questionnaire: • 23 out of 24 choose AV+ or G+ as first or second • 19 choose AV+ and G+ as the top two
More Qualitative Results • Participants liked both styles of results summaries • Participants disliked • Scrolling • Moving the mouse to see enhanced summaries • Hiding the url in enhanced summaries • Not seeing query terms in context (AV)
Quantitative Results • Task time • Artificial cutoff of 10 minutes assigned even if task not completed • Participants significantly faster with the enhanced summaries • There was a slight correlation between system and task completion, but not strong
Experiment Design Example: Marking Menus Based on Kurtenbach, Sellen, and Buxton, Some Articulartory and Cognitive Aspects of “Marking Menus”, Graphics Interface ‘94, http://reality.sgi.com/gordo_tor/papers
Experiment Design Example: Marking Menus • Pie marking menus can reveal • the available options • the relationship between mark and command • 1. User presses down with stylus • 2. Menu appears • 3. User marks the choice, an ink trail follows
Why Marking Menus? • Same movement for selecting command as for executing it • Supporting markings with pie menus should help transition between novice and expert • Useful for keyboardless devices • Useful for large screens • Pie menus have been shown to be faster than linear menus in certain situations
What do we want to know? • Are marking menus better than pie menus? • Do users have to see the menu? • Does leaving an “ink trail” make a difference? • Do people improve on these new menus as they practice? • Related questions: • What, if any, are the effects of different input devices? • What, if any, are the effects of different size menus?
Experiment Factors • Isolate the following factors (independent variables): • Menu condition • exposed, hidden, hidden w/marks (E,H,M) • Input device • mouse, stylus, track ball (M,S,T) • Number of items in menu • 4,5,7,8,11,12 (note: both odd and even) • Response variables (dependent variables): • Response Time • Number of Errors
Experiment Hypotheses • Note these are stated in terms of the factors (independent variables) • Exposed menus will yield faster response times and lower error rates, but not when menu size is small • Response variables will monotonically increase with menu size for exposed menus • Response time will be sensitive to number of menu choices for hidden menus (familiar ones will be easier, e.g., 8 and 12) • Stylus better than Mouse better than Track ball
Experiment Hypotheses • Device performance is independent of menu type • Performance on hidden menus (both marking and hidden) will improve steadily across trials. Performance on exposed menus will remain constant.
Experiment Design • Participants • 36 right-handed people • usually gender distribution is stated • considerable mouse experience • (almost) no trackball, stylus experience
Experiment Design • Task • Select target “slices” from a series of different pie menus as quickly and accurately as possible • (a) exposed (b) hidden • Can move mouse to select, as long as butten held down • Menus were simply numbered segments • (meaningful items would have longer learning times) • Participants saw running scores • Shown grayed-out feedback about which selected • Lose points for wrong selection
Experiment Design • 36 participants • One between-subjects factor • Menu View Type • Three levels: E, H, or M • (Exposed, Hidden, Marking) • Two within-subjects factors • Device Type • Three levels: M, T, or S • (Mouse, Trackball, Stylus) • Number of Menu Items • Six levels: 4, 5, 7, 8, 11, 12 • How should we arrange these?
Experiment Design E H M Between subjects design How to arrange the devices? 12 12 12
Experiment Design A Latin Square E H M M T S T S M No row or column share labels S M T 12 12 12 (Note: each of 12 participants does everything in one column)
Experiment Design Block by size then randomize the blocks. How to arrange the menu sizes? E H M M T S T S M S M T
Experiment Design Block by size then randomize the blocks. E H M M T S T S M S M T 5 11 12 8 (Note: the order of each set of menu size blocks will differ for each participant in each square) 7 4
5 7 8 11 12 12 5 8 4 7 11 4 Experiment Design E H M 40 trials per block M T S T S M S M T (Note: these blocks will look different for each participant.)
Experiment Overall Results So exposing menus is faster … or is it? Let’s factor things out more.
A Learning Effect When we graph over the number of trials, we find a difference between exposed and hidden menus. This suggests that participants may eventually become faster using marking menus (was hypothesized). A later study verified this.
Factoring to Expose Interactions • Increasing menu size increases selection time and number of errors (was hypothesized). • No differences across menu groups in terms of response time. • That is, until we factor by menu size AND menu group • Then we see that menu size has interaction effects on Hidden groups not seen in Exposed group • This was hypothesized (12 easier than 11)
Factoring to Expose Interactions • Stylus and mouse outperformed trackball (hypothesized) • Stylus and mouse the same (not hypothesized) • Initially, effect of input device did not interact with menu type • this is when comparing globally • BUT ... • More detailed analysis: • Compare both by menu type and device type • Stylus significantly faster with Marking group • Trackball significantly slower with Exposed group • Not hypothesized!
Average response time and errors as a function of device, menu size, and menu type. Potential explanations: Markings provide feedback for when stylus is pressed properly. Ink trail is consistent with the metaphor of using a pen.
E H M M T S T S M S M T Experiment Design How can we tell if order in which the device appears has an effect on the final outcome? Some evidence: There is no significant difference among devices in the Hidden group. Trackball was slowest and most error prone in all three cases. Still, there may be some hidden interactions, but unlikely to be strong given the previous graph.
Statistical Tests • Need to test for statistical significance • This is a big area • Assuming a normal distribution: • Students t-test to compare two variables • ANOVA to compare more than two variables
Summary • Formal studies can reveal detailed information but take extensive time/effort • Human participants entail special requirements • Experiment design involves • Factors, levels, participants, tasks, hypotheses • Important to consider which factors are likely to have real effects on the results, and isolate these • Analysis • Often need to involve a statistician to do it right • Need to determine statistical significance • Important to make plots and explore the data
Longitudinal Studies • Trace the use of an interface over time • Do people continue to use it, or drop it? • How does people’s use change over time?
Longitudinal Studies • Dumais et al. 2003 • Studied use of desktop search • Some people had sort by date as default, others had sort by relevance as default • A number of people switched from relevance to date; few went the other way • Kaki 2005 • Studied use of term-grouping search interface • People used the groups only for certain types of queries • People’s queries got shorter, since the interface could disambiguate for them.
Followup Work • Hierarchical Markup Menu study
Followup Work • Results of use of marking menus over an extended period of time • two person extended study • participants became much faster using gestures without viewing the menus
Followup Work • Results of use of marking menus over an extended period of time • participants temporarily returned to “novice” mode when they had been away from the system for a while
Wizard of Oz Studies • (discussed briefly in Nielsen) • Useful for simulating a smart program in order to get participant responses • Examples: Test out: • a speech interface • a question-answering interface There is a man behind the curtain!
Discuss Jeffries et al. • Compared 4 Evaluation Techniques • Heuristic Evaluation • Software Guidelines • Cognitive Walkthroughs • Usability Testing • Findings?