Astha Singh 1

Evgeny Chukharev 2

Mark Torrance 3

1 Department of Computer Science, Iowa State University, USA
2 Department of English, Iowa State University, USA
3 Psychology Department, Nottingham Trent University, UK

This material is based upon work supported by the US National Science Foundation under Grant No. 2302644.

Introduction

  • Spontaneous multi-sentence text production refers to tasks like writing an essay (term paper) in which writers need to generate / retrieve content as they write.
  • This is a semi-parallel, just-in-time process [1] that is typically performed with remarkable fluency.
  • Writers sometimes pause briefly at, for example, the start of sentences and look back into their already-written text. This typically involves short sequences of hopping back and forth among words and phrases, rather than sustained reading [2,3].
  • This behaviour does not seem consistent with error monitoring. It may, however, serve to support planning what to write next.
  • This would be consistent with the fact that writers who are prevented from looking back at their text pause for longer at sentence boundaries [4].
  • We present a first attempt at providing more direct evidence.

Methods

  • 30 undergraduates each wrote two short essays on non-specialist topics.
  • Keystrokes and eye-movements logged with CyWrite [2] allowing automatic extraction of fixated words within emerging text, and their containing sentences.
  • Human data comprised fixated words, text of current paragraph, and what they wrote next up to the next sentence terminator, for all instances where writer hesitated and looked back.
  • LLMs (presented results are just GPT-4) were prompted with the task assignment and the text of current paragraph and asked to provide a continuation either with or without “We have identified the following key words as particularly important: [fixated words]”.

Discussion

  • Greater overlap between meaning of participant and LLM completions when the LLM is prompted with the words (or sentences) that the participant fixated immediately before writing. This is consistent with fixated words cueing retrieval of what to say next.
  • Is this trivial? Possibly. We see it as a first step towards understanding how lookback is implicated in maintaining written-production fluency.
  • But LLMs might not process texts in the same way as humans?. True, but we don’t need to make that claim. Here LLM just provides a knowledge structure that is usefully similar to university students’.

References

1.
Roeser J, Conijn R, Chukharev E, Ofstad GH, Torrance M. Typing in tandem: Language planning in multisentence text production is fundamentally parallel. Journal of Experimental Psychology: General [Internet]. 2025 Jul;154(7):1824–54. Available from: https://doi.apa.org/doi/10.1037/xge0001759
2.
Chukharev-Hudilainen E, Saricaoglu A, Torrance M, Feng HH. Combined Deployable Keystroke Logging and Eyetracking for Investigating L2 Writing Fluency. Studies in Second Language Acquisition. 2019 Jul;41(3):583–604.
3.
Wengelin Å, Torrance M, Holmqvist K, Simpson S, Galbraith D, Johansson V, et al. Combined eyetracking and keystroke-logging methods for studying cognitive processes in text production. Behavior Research Methods. 2009 May;41(2):337–51.
4.
Torrance M, Rønneberg V, Johansson C, Uppstad PH. Adolescent Weak Decoders Writing in a Shallow Orthography: Process and Product. Scientific Studies of Reading. 2016 Sep;20(5):375–88.

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