Project Co-Ordinator

Horacio Saggion, Universitat Pompeu Fabra (UPF)

Principal Investigator at Leeds

Serge Sharoff

Co-Investigator at Leeds

Carlo Eugeni

Research Fellow at Leeds

Nouran Khallaf

Funding

EU Horizon

Time frame

2024-2026

Overall project aims

iDEM aims at addressing the linguistic barriers in deliberative and participatory democratic practices through a thorough intersectional analysis of conditions under which the existing structures of power limit participation of marginalised and vulnerable communities who might have limited skills in reading, writing or understanding a fairly complex language required for deliberative and participatory processes. Our vision in this project is to improve this situation by developing artificial intelligence and natural language processing models to make democratic spaces more inclusive and accessible for all. The democratic spaces to be proposed and implemented in realistic use cases will allow debate between different sectors of the society. This will be the first study to investigate artificial intelligence and natural language processing methods as enablers in addressing political inequality for deliberation and participation in democratic spaces.

Our specific aims in WP3

The University of Leeds leads WP3, Advanced Human Language Technology for non-discriminatory and inclusive democratic spaces

This WP will research, develop and deliver the iDEM text simplification system and services. This will include investigating mechanisms:

1. to identify complex linguistic phenomena in the specific context of the project,
2. to implement automatic transformation aiming at reducing the complexity of the information presentation making it more accessible to specific audiences,
3. to implement a text generation assistant in the content of deliberative democracy, and
4. to integrate the components into an open-API cloud-based web-services platform to deliver the iDEM services and Text-simplification App.

The WP will apply Deep Learning techniques to make the methods applicable to a range of different audiences, languages and domains. The solutions (algorithms and language resources) will be implemented as a freely available library for easy integration into Internet-based deliberative and participatory democratic spaces. In particular, we will investigate what makes texts and individual sentences difficult to specific audiences, how to make them more readable and how to detect whether this leads to distortions or hallucinations.

One of the key features of this research is the focus on explainability, i.e. understanding of why the language models prefer to simplify a sentence, thus ensuring that our AI models make their decisions for the right reason.

Tasks in Workpackage 3

T3.1

Text Difficulty Assessment and Classification Tool. M1-M30. Lead: UOLeeds. Participants: UPF, CAPITO

T3.2

Multilingual Adaptable Text Simplification System. M1-M30. Lead: UPF. Participants: UOLeeds, CAPITO

T3.3

Intrinsic and Expert Evaluation. M6-M31. Lead: CAPITO, Participants: UOLeeds, UPF, ANFFAS, PIM, IMPD

T3.4

Text Generation Assistant. M12-M34. Lead: UPF, Participants: PIM, ANFFAS, IMPD, UOLeeds

T3.5

Service Implementation. M6-M36. Lead: MAC. Participants: CAPITO, UOLeeds, UPF

Participants

Preliminary dataset

See the webpage with our multilingual dataset derived from Vikidia, as well as the tools used to make text simplification models.