Using Natural Language Processing (NLP) to assess changes in transdisciplinary understandings across a large research consortium

Researching how to create just and sustainable futures for all requires innovative and mission-oriented transdisciplinary approaches. This article sets out our approach and findings following the assessment, using Natural Language Processing (NLP), of the development in transdisciplinary understandings across a large and newly-forming research consortium as it changed over the six-year period of delivery (2019-2025). Two outcomes were sought from this assessment: i) learning to improve the transdisciplinary working in new, large research teams aiming to solve complex global challenges; and ii) improvement of use of NLP for this kind of exercise. A research-on-research workstream provided qualitative data from 63 semi-structured interviews carried out with 39 consortium members, with three rounds of interview being conducted in Years 2, 3 and 5 of the programme. Drawing on this data, a ‘dual approach’ to NLP was used to assess changes over time and by discipline: i) co-word clustering of full transcripts; and ii) analysis of vocabulary identified as transdisciplinary. The co-word clustering identified nine themes: new approaches (to operationalisation), case studies, the urban development context, mission-orientation, shared understandings, plural understandings, project structure and phasing, and collaboration. The analysis of transdisciplinary vocabulary indicated a clear convergence across disciplinary areas over six years and a small set of useful, jargon-free words (15 - 25% of 242 identified words were used across the four disciplinary areas). The results and analysis point to some useful findings and potential for improvement, but significant limitations were identified that would need addressing including: interpretation, thematic coherence, data quantum, inclusion of interviewer vocabulary, changes in programme development, normalisation and threshold determination. The approach was labour intensive, though future iterations should be much less so if the limitations can be addressed.