Slides: rpubs.com/jensroes/sig-paris-2024
Understanding factors that influence interkey intervals requires a theory of how the mental processes that underlie the generation of keystrokes are coordinated.
How do we produce texts?
Three general planning stages
Higher level info cascades into lower levels
Classic (serial) view of writing
Classic (serial) view of writing
Parallel view of writing (Olive 2014)
Parallel view of writing (Olive 2014)
Parallel view of writing (Olive 2014)
Serial view
- Pause duration before a writer starts a sentence is typically longer than before a mid-sentence word and these are longer than between mid-word key presses (e.g. Conijn, Roeser, and van Zaanen 2019).
- Long interkey intervals are typically followed by production bursts (Hayes 2012; Kaufer, Hayes, and Flower 1986).
- Planning message and syntax for a sentence adds to the time required to prepare the upcoming word, and the motor planning required to produce the first keystroke (Baaijen, Galbraith, and De Glopper 2012; Roeser, Torrance, and Baguley 2019).
Parallel view
- Variations in interkey intervals are not sufficiently explained by text location.
- Writers often do not hesitate before sentence (or word) onset.
- In some cases sentences are not planned, planning was postponed until after sentence onset, or completed in parallel with previous output (Nottbusch 2010; Roeser, Torrance, and Baguley 2019).
- Sentence initial durations are often shorter than what one would expect if sentence-initial pauses reflect planning.
- Fluent output is maintained by a cascade of processes that run in parallel (Olive 2014; Van Galen 1991).
How can we represent these models statistcally?
Serial view: single-process model
\[ \text{iki}_i \sim\text{ } \text{log}\mathcal{N}(\beta, \sigma_{e}^2) \]
Serial view: single-process model
\[ \text{iki}_i \sim\text{ } \text{log}\mathcal{N}(\beta_\text{textlocation[i]}, \sigma_{e}^2) \]
Parallel view can be implemented as two distributions mixture-process (Roeser et al. 2024)
- Executing motor movements (100-150 msecs, Conijn, Roeser, and van Zaanen 2019; Van Waes et al. 2021).
- If upstream processes provide output more slowly, then interkey intervals are determined by time taken to complete upstream processes and not motor movement.
Parallel view: mixture of two processes
\[ \text{iki}_{i} \sim\text{ } \theta \times \text{log}\mathcal{N}(\beta, \sigma_{e}^2) \\ \theta=1 \]
Parallel view: mixture of two processes
\[ \text{iki}_{i} \sim\text{ } \theta \times \text{log}\mathcal{N}(\beta + \delta, \sigma_{e'}^2) + \\ (1 - \theta) \times \text{log}\mathcal{N}(\beta, \sigma_{e}^2) \]
- Fluent typing speed: \(\beta\)
- Hesitation slowdown (pause duration): \(\delta\)
- Hesitation probability (pause frequency): \(\theta\)
Parallel view: mixture of two processes
\[ \text{iki}_{i} \sim\text{ } \theta_\text{location[i]} \times \text{log}\mathcal{N}(\beta + \delta_\text{location[i]}, \sigma_{e'_\text{location[i]}}^2) + \\ (1 - \theta_\text{location[i]}) \times \text{log}\mathcal{N}(\beta, \sigma_{e_\text{location[i]}}^2) \]
- Fluent typing speed: \(\beta\)
- Hesitation slowdown (pause duration): \(\delta\)
- Hesitation probability (pause frequency): \(\theta\)
How can we evaluate these models?
Six datasets with interkey intervals
| Dataset | Source | Keylogger | Task | N | Age | Sample | Country | Language |
|---|---|---|---|---|---|---|---|---|
| C2L1 | Rønneberg et al. (2022) | EyeWrite | Argumentative | 126 | 12 | 6th graders | Norway | Norwegian |
| CATO | Torrance et al. (2016) | EyeWrite | Expository | 52 | 17 | Secondary school students (dyslexic, non dyslexic) | Norway | Norwegian |
| GE2 | Ofstad Oxborough and Torrance (2011) | EyeWrite | Argumentative | 45 | 19 | Undergraduate students | UK | English |
| LIFT | Vandermeulen et al. (2020) | InputLog | Synthesis | 658 | 17 | Secondary school students | The Netherlands | Dutch |
| PLanTra | Rossetti and Van Waes (2022) | InputLog | Text simplification | 47 | 23 | Master students | Belgium | English (L2) |
| SPL2 | Torrance, Roeser, and Chukharev (n.d.) | CyWrite | Argumentative | 39 | 21 | Undergraduate students | USA | English |
Text location classifications
| Location | Example |
|---|---|
| Within word | T\(^{\wedge}\)h\(^{\wedge}\)e c\(^{\wedge}\)a\(^{\wedge}\)t m\(^{\wedge}\)e\(^{\wedge}\)o\(^{\wedge}\)w\(^{\wedge}\)e\(^{\wedge}\)d. T\(^{\wedge}\)h\(^{\wedge}\)e\(^{\wedge}\)n i\(^{\wedge}\)t s\(^{\wedge}\)l\(^{\wedge}\)e\(^{\wedge}\)p\(^{\wedge}\)t. |
| Below word | The \(^{\wedge}\)cat \(^{\wedge}\)meowed. Then \(^{\wedge}\)it \(^{\wedge}\)slept. |
| Before sentence | The cat meowed. \(^{\wedge}\)Then it slept. |
| a Note: Key intervals that terminated in a revision were removed. |
Model overview
| Models | Description |
|---|---|
| Serial | |
| M1 | Single distribution Gaussian |
| M2 | Single distribution log-Gaussian |
| M3 | Single distribution log-Gaussian with different variance components per text location |
| Parallel | |
| M4 | Two-distributions mixture of log-Gaussians |
Which model showed better performance?
Model comparisons
| Data set | Mixture process (M3 vs M4) | Single process (unequal var.; M2 vs M3) | Single process (M1 vs M2) |
|---|---|---|---|
| C2L1 | |||
| CATO | |||
| GE2 | |||
| LIFT | |||
| PLanTra | |||
| SPL2 |
Model comparisons
| Data set | Mixture process (M3 vs M4) | Single process (unequal var.; M2 vs M3) | Single process (M1 vs M2) |
|---|---|---|---|
| C2L1 | 23 | 13 | 47 |
| CATO | 24 | 18 | 43 |
| GE2 | 27 | 26 | 63 |
| LIFT | 40 | 21 | 46 |
| PLanTra | 26 | 17 | 64 |
| SPL2 | 18 | 25 | 69 |
Model fit to data
How do text locations affect the mixture process?
Mixture distributions
Location estimates by model parameter
| Dataset | before sentence vs word | before vs within word | before sentence vs word | before vs within word |
|---|---|---|---|---|
| C2L1 | 0.11 | -2.72 | -0.39 | 46.52 |
| CATO | 18.63 | -1.6 | -0.41 | 20.8 |
| GE2 | 33.53 | 3.29 | 18.31 | 26.33 |
| LIFT | -0.69 | -0.88 | 0.47 | 24.56 |
| PLanTra | 13.24 | 4.02 | -1.27 | 31.88 |
| SPL2 | 48.28 | 2.2 | 5.8 | 24.28 |
Location estimates by model parameter
Location estimates by model parameter
Planning text unfolds in parallel to production!
Planning text unfolds in parallel to production!
- First robust evidence that written composition is largely a parallel process.
- Writers do not necessarily pause at larger linguistic locations but plan utterances in parallel to writing.
- Evidenced by
- stronger predictive performance of mixture models (Roeser et al. 2024).
- pauses are not generally more likely before sentences compare to words (but often longer).
- Pauses are consistently more likely at word-initial location compared to word-medial but (1) not always longer and (2) for sentence-initial interkey intervals pausing behaviour varies as a function of writing experience, languages (e.g. young / L2 writers, students), composition tasks.
“Maybe mixture models are just always bettter?”
Simulation
- Simulate data from (i) a single log-normal distribution and (ii) a mixture of two log-normal distributions.
- Analysed both in (i) a single process model and (ii) a mixture model.
- Data simulated with mixture process:
- Advantage for mixture model over single process model: \(\Delta\widehat{elpd} =\) -191.3 (16.5)
- Data simulated with single process:
- Negligible difference between models: \(\Delta\widehat{elpd} =\) -0.5 (0.7)
Thank you!
We are grateful for all authors who made their available, in particular Nina Vandermeulen, Alessandra Rossetti, and colleagues.
This work was supported by
- National Science Foundation under Grant No. 2016868: “ProWrite: Biometric feedback for improving college students’ writing processes.” (Iowa State University)
- UKRI ESRC under Grant No. ES/W011832/1: “Can you use it in a sentence?: Establishing how word-production difficulties shape text formation.” (Nottingham Trent University)
For a tutorial on Bayesian mixture-model analysis in the context of keystroke data see https://rpubs.com/jensroes/mixture-models-tutorial.
References
Baaijen, Veerle M., David Galbraith, and Kees De Glopper. 2012. “Keystroke Analysis: Reflections on Procedures and Measures.” Written Communication 29 (3): 246–77.
Conijn, Rianne, Jens Roeser, and Menno van Zaanen. 2019. “Understanding the Keystroke Log: The Effect of Writing Task on Keystroke Features.” Reading and Writing 32 (9): 2353–74.
Hayes, John R. 2012. “Evidence from Language Bursts, Revision, and Transcription for Translation and Its Relation to Other Writing Processes.” In Translation of Thought to Written Text While Composing: Advancing Theory, Knowledge, Methods, and Applications, edited by M. Fayol, D. Alamargot, and V. Berninger, 15–25. Psychology Press.
Kaufer, David S., John R. Hayes, and Linda S. Flower. 1986. “Composing Written Sentences.” Research in the Teaching of English, 121–40.
Nottbusch, Guido. 2010. “Grammatical Planning, Execution, and Control in Written Sentence Production.” Reading and Writing 23 (7): 777–801.
Ofstad Oxborough, Gunn Helen, and Mark Torrance. 2011. “Multilevel Analysis of Latency in Writing.” In 21st Annual Meeting of the Society for Text and Discourse. http://textanddiscourse2011.conference.univ-poitiers.fr/PROG_DEFIN.pdf.
Olive, Thierry. 2014. “Toward a Parallel and Cascading Model of the Writing System: A Review of Research on Writing Processes Coordination.” Journal of Writing Research 6 (2): 173–94.
Roeser, Jens, Sven De Maeyer, Mariëlle Leijten, and Luuk Van Waes. 2024. “Modelling Typing Disfluencies as Finite Mixture Process.” Reading and Writing 37 (2): 359–84.
Roeser, Jens, Mark Torrance, and Thom Baguley. 2019. “Advance Planning in Written and Spoken Sentence Production.” Journal of Experimental Psychology: Learning, Memory, and Cognition 45 (11): 1983–2009. https://doi.org/10.1037/xlm0000685.
Rønneberg, Vibeke, Mark Torrance, Per Henning Uppstad, and Christer Johansson. 2022. “The Process-Disruption Hypothesis: How Spelling and Typing Skill Affects Written Composition Process and Product.” Psychological Research 86 (7): 2239–55.
Rossetti, Alessandra, and Luuk Van Waes. 2022. “Text Simplification in Second Language: Process and Product Data.” Zenodo. https://doi.org/10.5281/zenodo.6720290.
Sivula, Tuomas, Måns Magnusson, Asael Alonzo Matamoros, and Aki Vehtari. 2020. “Uncertainty in Bayesian Leave-One-Out Cross-Validation Based Model Comparison.” arXiv Preprint arXiv:2008.10296.
Torrance, Mark, Jens Roeser, and Evgeny Chukharev. n.d. “Lookback Supports Cascaded, Just-in-Time Processing in Second Language Written Composition.”
Torrance, Mark, Vibeke Rønneberg, Christer Johansson, and Per Henning Uppstad. 2016. “Adolescent Weak Decoders Writing in a Shallow Orthography: Process and Product.” Scientific Studies of Reading 20 (5): 375–88.
Van Galen, Gerard P. 1991. “Handwriting: Issues for a Psychomotor Theory.” Human Movement Science 10 (2): 165–91.
Van Waes, Luuk, Mariëlle Leijten, Jens Roeser, Thierry Olive, and Joachim Grabowski. 2021. “Measuring and Assessing Typing Skills in Writing Research.” Journal of Writing Research 13 (1): 107–53. https://doi.org/10.17239/jowr-2021.13.01.04.
Vandermeulen, Nina, Sven De Maeyer, Elke Van Steendam, Marije Lesterhuis, Huub Van den Bergh, and Gert Rijlaarsdam. 2020. “Mapping Synthesis Writing in Various Levels of Dutch Upper-Secondary Education: A National Baseline Study on Text Quality, Writing Process and Students’ Perspectives on Writing.” Pedagogische Studiën: Tijdschrift Voor Onderwijskunde En Opvoedkunde 97 (3): 187–236.
Vehtari, Aki, Andrew Gelman, and Jonah Gabry. 2015. “Pareto Smoothed Importance Sampling.” arXiv Preprint arXiv:1507.02646.
———. 2017. “Practical Bayesian Model Evaluation Using Leave-One-Out Cross-Validation and WAIC.” Statistics and Computing 27 (5): 1413–32.