Early attempts late XIX c.; early 1950’s - using a movie camera and hand-coding (Fitts, Jones & Milton 1950)Now computerized and “easy to use” infrared light sources and camerasstationary and mobile
Eye tracking can allow identification of the specific content acquired by the person from Web pages Eye tracking enables high resolution analysis of users’ activity during interactions with information systemsEye-tracking in HCI/Usability applications have frequently analyzed eye-movement and eye-gazeposition aggregates ('hot spots’)spatiotemporal-intensity – attentional maps (heat maps)also sequential – scan paths
eye gaze points eye gaze points in screen coordinates + distance eye fixations in screen coordinates + validity pupil diameter [head position 3D, distance from monitor]50/60Hz; 300Hz; 1000-2000Hz eye-trackerscommon in the USA: 60Hz: one data record every 16.67ms; in EU: 50Hz 20msHigher-order patterns:! reading models & derived measures
Modeling reading eye movement to investigate (textual) information seekingWe have developed a methodology to analyze eye-movement patterns:Model the reading process to represent (textual) information acquisition in searchMeasure the cognitive effort due to (textual) information acquisitionUse both to correlate higher-level constructs (task characteristics, user domain knowledge, etc.)
Eye-mind link hypothesis: attention is where eyes are focused (Just & Carpenter, 1980; 1987)Attention can move with no eye movement BUT eyes cannot move without attentionCombined top-down and bottom-up control language processing (higher-level) controls when eyes move, while visual processing (lower-level) controls where eyes move (Reichle et al., 1998)Eye movements are cognitively controlled (Findlay & Gilchrist, 2003)Eyes fixate until cognitive processing is completed (Rayner, 1998)Eye gaze pattern analysis is powerful:Eye gaze is only way to acquire (textual) information1. + 2. Direct causal connection between observable information acquisition behavior (eye-tracker) and user’s mental state
Eye tracking work on reading behavior in information search have mostly analyzed eye gaze position aggregates ('hot spots').This does not address the fixation sub-sequences that are true reading behavior.
Our algorithm distinguishes readingfixation sequences from isolated fixations, called 'scanning' fixations. Filter: Minimum time required for lexical processing: fixation duration ( > 150ms)(Pollatsek et al, 2006; Reichle et al, 2006) (min of mean time for lexical processing completion)Transition probabilities between statesEach state characterized by: number of lexical fixations,durationlength of eye-movement trajectory (scan path) amount of text covered
Reading speed will be slower for:hard to read text (Rayner & Pollatsek, 1989); more complex concepts (Morris, 1994)unfamiliar words (Williams & Morris, 2004); words used in less frequent senses (Sereno, O’Donnell, & Rayner, 2006)Perceptual span reflects a human limitation on the number and difficulty of concepts that can be processed (e.g. Pollatsek et al. 1986)Regressions:10-15% of fixations are regressions; Reading goal affects reading regressionsMore regressions when: conceptually complex & difficult text passages, resolution of ambiguous (sense) wordsGreater LFDE (fixation duration) indicates less familiar words & greater conceptual complexityFDis also correlated with establishing word meaning in context.
Tasks varied in several dimensions: (mention Systematic Review of Imposed Search Tasks)complexity defined as the number of necessary steps needed to achieve the task goal (e.g. identifying an expert and then finding their contact information), the task product (factual vs. intellectual, e.g. fact checking vs. production of a document), the information object (a complete document vs. a document segment), andthe nature of the task goal (specific vs. amorphous).
Task difficulty was self-rated by participants after each search was completed(7-point Likert scale: ’very easy’ to ’extremely difficult’)Search effort: task time, pages visited, queries enteredCopy Editing (CPE) required the least effort of all tasksAdvance Obituary (OBI) required overall most effortCognitive effort measures: greater perceived difficulty(self-ratings) and higher search task effort correlated with higher median LFDE (Kruskal-Wallis χ2 =12.5, p<0.05) slower reading speed (ANOVA F=5.5 p<0.05)Comparing all tasks, all pages, page levelfor all tasks - sig diffs: mean perceptual span (K-W, χ2 =11.5, p<.01), mean reading speed (ANOVA, F=8.3, p<.01), mean fixation duration (K-W, χ2 =11.5, p<.01), total regressions (K-W, χ2 =7.6, p=.053)e.g. for CPE: longer perceptual span, faster, less regressions, higher LFDECognitive effort measures seem validEye gaze pattern cognitive effort measures match with subjective task difficultyCognitive effort measure results correlate with task characteristics related to task efforte.g. Complex tasks, amorphous goals
Level - the information object to process (a complete document vs. a document segment)Task Goal - the nature of the task goal (specific vs. amorphous)Searchers are adopting different reading strategies for different task types
Participants’ domain knowledge (PDK) was measured by a sum of term ratings participants rated 409 MeSH(Medical Subject Headings) terms (out of 25186)normalized by a hypothetical expertthe terms were the terms in three MeSH trees (each one a MeSH category -- Genetic processes (G05), genetic phenomena (G13), and genetic structures (G14)the anchors in the MeSH rating provide clean points for participant self ratings 'no knowledge' and 'can explain to others' are bright lines. 1 was 'vague idea' and 4 was 'high knowledge' and designed to be constrasted against the anchors. 3 was 'some knowledge' and that is quite mushy.
RF are an ensemble learning technique invented by Leo Breiman in 2001. their advantages include:
RF are an ensemble learning technique invented by Leo Breiman in 2001.
--mention the Tobii laptop prototype as a way of showing the idea of deployed eye tracking systems is not crazy.
There is more to eye tracking data then eye-movement locations with timestamps goal to : infer and predict context and mental states of a person engaged in interactive information seeking. use of measures ; eye-movement patterns and interaction logs to infer dynamic user states (such as cognitive load), task characteristics, persistent user characteristics (such as domain knowledge)
… a note on eye-tracking methodology : If we really can predict level of knowledge what can be done with that (potentially) to improve systems? eye movement analysis can be conducted in near real time – online (only a few seconds of data are needed). Predictions for new data using previously learned models (for example the RF models) is instantaneous. This enables actions by the system at the levels of just a few seconds.Eye tracking enables high resolution analysis of searcher’s activity during interactions with information systemsThere is more beyond eye-gaze locations with timestamps Eye-tracking data:can support for identification of search task typesreflects differences in searcher performance on user interfacesreflects individual differences between searchersHigh potential for implicit detection of a searcher’s states
Towards Interaction Models Derived From Eye-tracking Data .
Towards Interaction Models Derived From Eye-tracking Data . Jacek Gwizdka & Michael Cole Rutgers University, USA email@example.com http://jsg.tel April 19, 2012
There is a Lot More Eye-tracking Data can offer UX / HCI / IA 4
Eye-tracking Data Patterns State1 State2 State3 5
Eye-movement Patterns New methodology to analyze eye-movement patterns ◦ Model reading and Measure cognitive effort ◦ Correlate with higher-level constructs user task characteristics, user knowledge, etc. 6
Reading Model Origins Based on E-Z Reader modelRayner , Pollatsek, Reichle ◦ Serial reading ◦ Words can be identified in parafovialregion ◦ Early lexical access (word familiarity) + Complete lexical processing (word identification) 2o (70px) foveal region parafoveal regiona bit MORE… 8
Two-State Reading Model q isolated fixation fixations sequences p Read Scan 1-p 1-q ◦ Filter fixations <150ms (min time required for lexical processing) ◦ Model states characterized by: probability of transitions; number of lexical fixations; duration length of eye-movement trajectory, amount of text covereda bit MORE… 9
Example Reading SequenceReading sequence:Fixation model states: (F F F) S (F R F) S SSSF (F F R F F F) F R F F FFF F S Reading state – R | Scanning state – S 10
Cognitive Effort Measures of Reading Reading Speed foveal region regression Fixation Regression a b c d Perceptual Span Perceptual span = Mean(a,b,c,d) Fixation Duration excess (“lexical processing excess”) 11
User Study 1: Cognitive Effort and Tasks N=32 Journalists’ Information Search OBI: advanced obituary INT: interview preparation CPE: copy editing BIC: background informationMORE… 12
Eye-data and Cognitive Effort Measures Subjective Task DifficultyTask complexity Cognitive effort measures reading speedby design mean fixation duration Copy Editing (CPE) perceptual span Advance Obituary (OBI) total fixation regressions CPEINTBIC OBI As expected: Copy Editing CPE easiest Search effort Advance Obituary OBI most difficult task time Sig: Kruskal-Wallis χ2 =46.1, p<.0001 pages visited queries entered 13
Eye-data and Task Characteristics q p Read Scan 1-p 1-q Interview preparation Copy Editing Measure Related Task Characteristics Frequencyo SRbias to read Advanced obituary and Interview preparationtasks: search f reading for document; task goalnot specific state Copy Editingtask: search for segment and task goal specific transitions RS bias to scanMORE… 14
User Study 2: Assessing User’s Knowledge Search in Genomics Domain N=40 Rate own domain knowledgeMORE… 15
Results: Modeling Domain Knowledge Eye-tracking Data Reading Model features & cognitive effort measures self- rated Domain knowledge MeSH-based self-ratings predicted 16
Results: Modeling Domain KnowledgeEye-tracking Data Reading Model features & build model Random cognitive effort measures Forest Model reading seq length and total duration For perceptual spanReading Model each fixation duration user predict regressions… self- rated Domain knowledge MeSH-based self-ratings agglomerative m hierarchical (k i * t i ) clustering PDK = i 1 (Ward’s) 5*m PDK: Participants’ domain knowledge predictedMORE… 17