Brief lookback cues content generation in spontaneous multi-sentence text production

Mark Torrance

Nottingham Trent University

Astha Singh

Iowa State University

Evgeny Chukharev

Iowa State University

background

  • Text composition is a semi-parallel, just-in-time process Roeser et al. (2025)

  • When writers hesitate at, for example, sentence boundaries, they frequently look back into the text that they have already written.

  • This typically involves hopping between words rather than sustained reading. Chukharev-Hudilainen et al. (2019), Torrance, Johansson, et al. (2016), Wengelin et al. (2009)

Hypothesis: Lookback supports planning what to write next.

(weak) evidence so far

Hypothesis: Lookback supports planning what to write next.

  • Mid-sentence lookback, without error correction, is more frequent and extensive when writing in non-fluent second language Chukharev et al. (2025)
  • Masking text-already-written (XX XXXX, XXXX X XXX. XXX) increases hesitation at sentence boundaries, particularly in students with dyslexia Torrance, Rønneberg, et al. (2016)

data

Eye movement and keystrokes from 30 university student writers composing short argumentative essays

  • Inter-keystroke intervals → identify pauses in production
  • Eye movement → which words were fixated when writers looked back during pauses

Using CyWrite
Chukharev-Hudilainen et al. (2019), https://github.com/chukharev/cywrite

analysis

Participant completion text compared with LLM (GPT-4) generated completion text in each of four prompting conditions:

  1. pretext only (task statement and already-written text; control)
  2. pretext + fixated words
  3. pretext + sentences containing fixated words
  4. pretext + matched non-fixated words (control)

Dependent variable: Semantic overlap between participant and LLM generated completion texts (cosine from OpenAI text-embedding-ada-002)

results

LLM-generated text aligns more closely in meaning with participant text when the model is prompted using the words that the participant fixated immediately prior

very provisional conclusion

First direct evidence that looking back into text-already-written cues / reinstates message for upcoming text.

slides | paper | code | references | funder

Slides: https://rpubs.com/mark-torrance/EPS_spring2026

Paper: https://aclanthology.org/2025.bea-1.61.pdf

Code: https://go.chukharev.com/bea-2025]

References

Chukharev, E., Roeser, J., & Torrance, M. (2025). Lookback supports semi-parallel, just-in-time processing in second language written composition. PLOS One, 20(11), e0334960. https://doi.org/10.1371/journal.pone.0334960
Chukharev-Hudilainen, E., Saricaoglu, A., Torrance, M., & Feng, H.-H. (2019). Combined Deployable Keystroke Logging and Eyetracking for Investigating L2 Writing Fluency. Studies in Second Language Acquisition, 41(3), 583–604. https://doi.org/10.1017/S027226311900007X
Roeser, J., Conijn, R., Chukharev, E., Ofstad, G. H., & Torrance, M. (2025). Typing in tandem: Language planning in multisentence text production is fundamentally parallel. Journal of Experimental Psychology: General, 154(7), 1824–1854. https://doi.org/10.1037/xge0001759
Torrance, M., Johansson, R., Johansson, V., & Wengelin, Å. (2016). Reading during the composition of multi-sentence texts: an eye-movement study. Psychological Research, 80(5), 729–743. https://doi.org/10.1007/s00426-015-0683-8
Torrance, M., Rønneberg, V., Johansson, C., & Uppstad, P. H. (2016). Adolescent Weak Decoders Writing in a Shallow Orthography: Process and Product. Scientific Studies of Reading, 20(5), 375–388. https://doi.org/10.1080/10888438.2016.1205071
Wengelin, Å., Torrance, M., Holmqvist, K., Simpson, S., Galbraith, D., Johansson, V., & Johansson, R. (2009). Combined eyetracking and keystroke-logging methods for studying cognitive processes in text production. Behavior Research Methods, 41(2), 337–351. https://doi.org/10.3758/BRM.41.2.337

Funding: This material is based upon work supported by the National Science Foundation under Grants No. 2016868 and 2302644.