Creative Writing and Dynamic Analysis of Writers' Manuscripts – CreaLAME

Cré@lame proposes a study of creative writing from the point of view of genetic operations and writing processes, their psycholinguistic dimension and the ergonomics of creativity. The hypothesis is that these operations can be modelled and analysed in the practice of a heritage writer, whose writing operations are preserved in drafts, or of a writer recorded by typing on a keyboard. To this end, Cré@LAME aims to develop digital tools for :
-capture and analyse the graphic traces of a succession of drafts by integrating, making visible and marking out deletions, additions, moves and replacements
-extraction of the lexical, morphosyntactic and genetic data thus created for as complete a writing sequence as possible
-modelling and summarising the qualified data to provide an interdisciplinary analysis (digital humanities, text genetics, literary studies, psycholinguistics)
-visual reconstruction and commentary of the writing film.
These tools will provide a direct and intelligible approach to the draft for the purposes of learning and literary mediation, and will also offer a dashboard of the writing in progress and thus of other reading and writing practices, likely to significantly modify uses and practices.
These tools will extend those developed in genetic criticism (MEDITE, Genographix) and complement those used in psycholinguistics (Inputlog, Sciptlog). They will add to the corpora usually studied data extracted from writers' drafts and live recordings in order to broaden the scope of research and provide a very high level term of comparison for the analysis of textualisation processes.
These data will make it possible to identify whether the operations are more or less identical from one writer to another, and whether, finally, constants emerge alongside regular practices specific to a given author, which is essential for the didactics of writing and, in particular, creative writing.
These elements will be compared with data of the same type from the analysis of successive versions of Wikipedia pages, a sample of collective informational writing that will provide a basis for comparison with writers' practices.
In the field of AI, the question will be whether these models can be transferred to the automatic generation of texts. The project focuses on the development of generative neural architectures based on models used in systems such as ChatGPT, Gemini and Copilot. In our case, the input sequence is one version of a text and the output sequence is another version of the same text. We will assess the extent to which existing architectures allow such a task to be carried out, and at what level of granularity. On this point, the study of prompts based on human-machine interaction will be important.
We will also be looking at the nature of the training data, which usually consists of single texts or pairs of static texts. It will be possible to learn evolving data and thus refine a large existing language model. Several studies confirm the nature and degree of impact of the weights of the final and intermediate representations. These are associated with identifiable linguistic notions, and their manipulation makes it possible to take more or less account of the neighbourhood or to obstruct learning biases. One question is whether this observation is valid in our case, which raises the question of the explicability of generative approaches and the identification of biases. It would then no longer be a matter of simply predicting or suggesting the next word, but of generating a list of words to avoid because of their triviality, for example, in order to move towards the more creative functions of generative AI.