Single Paper Session 2

Session Page

Session Information

This page shows the session details and the presentations assigned to this session.

Back to programme overview
Add this session to calendar

Literary Writing Process Modeling: across manuscript drafts and digital traces

Keywords
AI writing;creative writing;keystroke logging data processing and analysis;writing models
Domain
Learning and Instructional Technology

Abstract

Investigating literary writing dynamics and authors’ revision signatures is increasingly recognized as a crucial field, drawing on both genetic criticism and psycholinguistics, as well as advanced generative AI systems. Despite this growing interest, a combined analysis of heritage manuscripts alongside contemporary keystroke logging data remains largely uncharted. Therefore, this proposal aims to bridge this gap by proposing a fine-grained modeling of literary writing and revision processes, developed within the Cré@LAME project (Literary Cre@tion and Author Manuscript Analysis), supporting an interactive assisted rewriting system, attuned to the author’s profile and revision strategies.The approach relies on a set of LLM-based agents specialized in context-aware rewriting, each performing a specific editorial role aligned with distinct revision intentions. These agents are coordinated by a multi-layer, multi-view Graph Neural Network (GNN) that models the evolution of textual states across heterogeneous materials, from linear manuscript transcriptions to digital writing traces.This network captures both linguistic (lexical, syntactic, semantic) and revision-oriented dimensions, reflecting editing operations and authorial intentions, across multiple levels, while guiding the agents’ rewriting operations according to learned patterns of textual evolution. This GNN thereby maintains coherence in editing operations while tracking author-specific revision practices.Accordingly, this work introduces a novel computational framework for textual genesis that addresses key aspects, including multi-granular data heterogeneity across manuscripts and digital log files, the inference of relevant indicators of authorial revision trajectories, and unified hierarchical representations formats of revision processes, integrating cross-source materials, suitable for multi-level graph modeling.Overall, this contribution advances research on textual genesis by highlighting how the integrated modeling of manuscript materials and digital traces provides deeper insights into authorial practices and the dynamics of literary creation.

Teenagers writing expository texts with and without gen-AI

Keywords
AI writing;writing models;writing process data collections;written language competence
Domain
Teaching and Teacher Education

Abstract

Writing with generative AI (gen-AI) introduces new affordances and constraints that invite reconsideration of long-standing writing models, such as Hayes and Flower’s (1980) framework and Kellogg’s (1996) model of working memory in writing. This study examines how key writing processes—planning, translating, and revising—and related subprocesses such as goal setting, audience adaptation, reading, and evaluation unfold when students write with gen-AI. Adopting a developmental perspective, it qualitatively compares writing with and without AI support in a cross-sectional design involving students (n = 52) aged 13, 15, and 18. In a classroom setting, the students produced comparable expository texts first without and then with a gen-AI tool of their choice, while their writing was captured through screen recordings. Post-task interviews probed their strategies and reflections on differences between the modes.These questions guide the study: (1) How do writing processes unfold with and without using gen-AI, and are there age-related differences? (2) How do writers interact with the gen-AI tool (e.g., prompting), and how do they make use of the generated text?Initial results show that all students produce coherent and linguistically appropriate expository texts without AI, consistent with earlier descriptions of developmental writing (e.g., Johansson, 2009; Wengelin et al., 2014). In contrast, age-related patterns emerge when AI is introduced. The 13-year-olds often use gen-AI to produce full texts based on task prompts and report valuing the tool’s ability to generate lengthy responses. The 15-year-olds tend to use gen-AI primarily for idea generation, rewriting the AI-generated material to align with their own voice. The 18-year-olds more often use gen-AI to refine their existing ideas and strengthen the logic of their texts.This developmental trend demonstrate that the youngest writers rely on gen-AI mainly to support translating processes, the middle group for planning processes, and the oldest group for revising processes. The findings are discussed in relation to how gen-AI may differentially support components of the writing process depending on writers’ developmental needs and strategic awareness, and how the use of gen-AI during writing can reshape existing writing models. Understanding these evolving practices is also essential for instructional approaches including assessment.

What Can Sentence-Centric Writing Models Reveal about the Writing Process?

Keywords
keystroke logging data processing and analysis;natural language processing (NLP);writing and linguistics;writing models
Domain
-

Abstract

Sentences are fundamental communicative units (Bühler 1918), and written texts are generally understood to consist of these units, but research on how writers produce sentences remains limited. Although linguistic modeling of the writing process has gathered interest in recent years, existing approaches, whether grounded in linguistic theory or in writing research, remain insufficient to explain how writers actually produce and revise text at a linguistic level. Prior work has investigated correspondences between writing bursts and linguistic structure (e.g., Kaufer et al. 1986; Cislaru and Olive 2018; Feltgen et al. 2023), examined revisions from a linguistic perspective (e.g., Manseri and Jouvenel 2025), proposed methods for transforming writing-process data into linguistic units (e.g., Leijten et al. 2019), and provided initial contributions to sentence-level analyses of the writing process (Miletic et al. 2022; Mahlow et al. 2024; Ulasik and Miletic 2024).We advance the state of the art by building on these developments and on the theoretical framework for sentence-centric modeling introduced by Ulasik et al. (2025). Our approach enables detailed tracking of sentence production through the analysis of sentence transformations and the detection of pauses at sentence boundaries. It supports systematic identification of bursts within sentence production and offers a method for characterizing the scope of transformations and bursts with respect to individual sentences.To investigate the potential of the model, we apply our software tool for sentence-centric modeling of writing, THEtool (https://github.com/mulasik/wta), to real-world data from the KLiCKe corpus (Yu Tian et al. 2025). This demonstrates the potential insights that emerge when shifting the analytical perspective from bursts or revisions to a sentence-centric view.