Deliberative and participatory processes currently don't have full legitimacy due to the exclusion and marginalisation of several vulnerable communities from democratic spaces. It is well documented that people with limited language skills, such as people with cognitive disabilities, struggle to participate in democratic processes. This happens in spite of the advocacy work of organizations which promote human rights. The iDEM project aims at addressing the barriers in deliberative and participatory democratic practices through a thorough intersectional analysis of conditions under which current structures and systems limit participation of marginalised and vulnerable communities. Specially those with limited skills in reading, writing or understanding a fairly complex language, which is often required for deliberative and participatory processes. iDEM will lay the theoretical foundations for the analysis of current marginalisation from deliberative processes of diverse under-represented groups due to a lack of language skills. It will adopt a user-centred approach for making participatory processes more accessible and inclusive, developing advanced natural language processing technologies (NLP) and artificial intelligence (AI) to empower under-represented groups, with tools to facilitate communication and dialog in democratic spaces. iDEM will co-create the next-generation multilingual models aimed at: (1) detecting possible sources of problems in understanding messages for several European languages and audiences, (2) automatically adapting texts in those languages to be accessible and unbiased for these audiences, (3) providing AI tools for enhancing the generation of appropriate messages and discourses. iDEM aims to create more accessible democratic spaces in Italy and Spain with customised technology enhancing the participation and representation of marginalised groups by providing unbiased and inclusive technology.
5,
civic participation, no matter how inclusive deliberative and participatory processes and platforms are, people with language dificulties are left behind. The proper dynamics of participation and the language used by policy makers and institutions could exacerbate, rather than improve, the under representation of such marginalised groups. People must be able to understand the information and deliberate on an equal basis, which imposes obligations on institutions. The United Nations Convention on the Rights of Persons with Disabilities (CRPD) includes accessibility as one of its rights 1. That is the ability of any product, service, content, environment, etc. to be used by people with the widest range of abilities. The CRPD also considers accessibility as an enabler for democratic participation rights, e.g., freedom of expression and self-determination. Consequently,
ments/convention-rights-persons-disabilities a lack of accessibility in democratic spaces can be 1. Examples are taken from the UK government voting linked to a risk of exclusion for persons who cannot information4 5, the Spanish Constitution 6 7, the Catalan participate equally. Against this background we have Parliament 8 9, and the European Union portal 10 11 . proposed the iDEM project “Innovative and Inclusive From a computational viewpoint, transforming complex Democratic Spaces for Deliberation and Participation” to texts into easy to read and understand ones, has already address the accessibility limitations in democratic spaces. been addressed in the field of automatic text simplification [5] which has usually concentrated on two diferent tasks: lexical simplification and syntactic simplification , 1.1. Context: Information Access each addressing diferent sub-problems. Lexical simplification will attempt to modify the vocabulary (e.g. target Literacy is fundamental to human development as it en- complex words) by choosing substitutes which are more ables people to contribute to their communities and to appropriate for the reader [6]. Changing words in context society. People who encounter dificulties making mean- is not an easy task because it may alter the meaning of ing out of content are a diverse group of individuals the original text. Syntactic simplification will transform with varying ranges of reading, writing, and understand- complex sentences into more readable or understandable ing abilities. This part of the population includes, for equivalents. For example, relative or subordinate clauses instance, those with low levels of literacy, intellectual or passive constructions, which may be more dificult to disabilities, dyslexia, aphasia, temporary impairments, or understand, could be transformed into simpler sentences with limited language skills (e.g. second-language learn- or into active form. Although some non-English research ers, immigrants, and displaced populations). Finally, the has been produced in this area, it is fair to say most reelderly can also fall into this category, as they can be search and resources have been generated for the English afected by a significant cognitive, physical, or sensory language. Our project contributes with solutions also for decline with age. Catalan, Italian, and Spanish.
Challenges of including people with disabilities in deliberative or participatory processes are even bigger than those of including them in representative processes. An 2. The iDEM Project estimated 800,000 EU citizens from 16 Member States were not able to participate in the 2019 European Parliament elections due to barriers such as limited accessibility to polling stations or insuficiently accessible information about candidates and debates2. But the number of people excluded from deliberative processes due to linguistic barriers, the focus of this project, is exponentially higher if we consider that 1% of the human population is afected by an intellectual disability 3.
Italian. Data will be gathered from past participatory processes, democratic institutions, and the Web. The annotated corpus will be used to fine-tune natural language processing models to automatically identify and classify the source of text complexity and simplifying them accordingly. We will also assist our users with tools for the generation of coherent discourses adopting carefully ifne-tuned Large Language Models. By working with associations for people with intellectual disabilities, iDEM will adopt a user-centred approach in use case design and corpus creation to ensure maximum impact in the community thus contributing to make democracy more accessible and inclusive. An iDEM innovative service will be created to deploy the developed language tech- Figure 1: Illustration of the iDEM Concept. nologies: it will be open-source, and well documented. It will provide an architecture for communication, as well as the open iDEM API. More concretely, iDEM objectives are: • Understand the limitations of current deliberative and participatory democratic practices in terms of language comprehension and production and propose accessible and innovative solutions to remove inequalities to make democratic processes widely accessible • Design a solution for making deliberative and participatory processes, at diferent levels of government, more inclusive, specially for marginalised and vulnerable communities • Implement and evaluate state of the art natural language processing technology in the areas of text accessibility and argumentative discourse generation for deliberative processes to facilitate engagement and enhance participation of otherwise marginalised communities • Pilot the solution with real users to evaluate the level of inclusiveness and accessibility of the novel proposed democratic spaces
citizens’ representatives, non-profit organisations and software developers which form an interdisciplinary and complementary team from five EU countries (Austria, Germany, Ireland, Italy, and Spain) and one non-EU country (UK).
methods for syntactic simplification and corpus-based 1. Universitat Pompeu Fabra (UPF, Spain): (a) Dept. of unsupervised techniques for lexical simplifications [ 8].
Information & Communication Technologies (DTIC) Where parallel complex-simple sentences are available and (b) Faculty of Law (e.g. English Wikipedia and Simple English Wikipedia 2. University of Leeds (UOL, UK) pairs) text simplification can be addressed as monolin3. CFS GmbH (CAPITO, Austria) gual Statistical Machine Translation (SMT) [9, 10]. In 4. INnEteXrdUiSszIinplsitnitauret FfuorrsKchoionpgeGraMtiBonHs(NMEaXnaUgSe,mGeernmt annuyd) recent years, Neural Machine Translation based system 5. Organización de entidades en favor de las personas con (NMT) have been applied in text simplification [ 11], outDiscapacidad Intelectual de la Comunidad de Madrid performing SMT models in terms of simplicity, adequacy, (Plena Inclusión Madrid) (PIM, Spain) and content preservation of the output. After the release 6. The National Microelectronics Applications Centre of the Transformer architecture in NLP [12], an approach
Ltd (MAC, Ireland) based on it was applied to simplification, obtaining state 7. Anfas Nazionale (ANFFAS, Italy) of the art performance by including, in addition to the par8. Fundación Cibervoluntarios (CIB, Spain) allel complex-simple training data, a paraphrase database. 9. ActionAid International Italia E.T.S. (AAIT, Italy) More recently, there has been increased interest in condi10. Institut Municipal de Persones amb Discapacitat tional training sequence-to-sequence models (e.g. embed11. (BIMarPceDlo,nSpaaOinm) budsman’s Ofice (BOO, Spain) ding control tokens during training), being applied for example to control summarization output or to control
The consortium’s interdisciplinary expertise enables the style of the generated text (e.g. politeness). In text to achieve the iDEM objectives: simplification, this method was used to control the “grade
UPF covers two key areas of research: theory of delib- level” of the simplified text, its length, or its syntactic erative democracy and new technologies at the service of complexity, achieving very promising results. Recent citizen participation, and from a technological perspec- research has shown that adding control tokens [13] does tive, experts in automatic text simplification. NEXUS help improve the performance of sentence simplificabrings experience in the design of deliberative and partic- tion models quite significantly, achieving state of the art ipative processes at national and international levels ap- in English and Spanish [14] using large pre-trained lanplying user-centred design to promote collaboration and guage models [15] (e.g. T5/mT5 from Google or mBART sustainable consensus in designing our solution, from a from Facebook). Concerning lexical simplification, sevpractical social science perspective, CAPITO brings its in- eral past approaches used traditional raw count vectors dustrial expertise in text simplification and easy-to-read word-vectors and available dictionaries for modelling annotation technology, involving users in creating and word semantics and to select simple word replacement for annotating corpora for natural language processing. UOL complex words [16]; these traditional vector models have has substantial expertise in translation, text classification, recently been substituted in simplification approaches and developing algorithms for low-resources domains by word embedding, which are learned from huge text and languages. MAC, an Irish innovation company, will collections. Nowadays, large-scale language models such deliver leading-edge products and services in software, as BERT and its variations have been applied to predict electronics and wireless communications to implement substitution candidates for complex words. For examthe solution with the technical experts. User organisa- ple, LS-BERT uses the masked language model (MLM) tions engage users: PIM federates around 120 associa- of BERT to predict a set of candidate substitution words tions and foundations that work with our target group, and their associated probability [17, 18]. In this context, ANFFAS federates around 200 associations and founda- the MLM predicts substitute words which are ranked tions for people with disabilities, while IMPD works for for simplicity using: probabilities, a language model, a the rights of the persons with disability to be included paraphrase database, word frequency and word semanin all aspects of society in the City of Barcelona. CIB tic similarity with the target word. These models, howin Spain and AAIT in Italy, are ONGs working in social ever, have clearly neglected bias-related aspects when issues engaging the civil society in civic participation, proposing simplifications. Generative AI, such as Chatresiliency, responsibility and democratic quality, which GPT – based on Generative Pre-trained Transformers are key to engage participation and drive testing of the (GPT) [19], show the possibilities of In-Context Learning. solution with citizens. Finally, BOO brings the human However, at the moment they have been outperformed rights perspective by engaging underrepresented citizens by specialised training. in democratic processes.
E-democracy refers to the use of information and commu- develop and implement three use cases each concerned nications technology (ICT) to create channels for public with a diferent language (Catalan, Italian or Spanish), consultation and participation, for example for elections, intersectionality (people with disabilities, migrants, the consultations or referendums, and over the years several elderly), and deliberative approach (citizen’s assembly, online platforms have been introduced. These platforms mini-public, consultation). The use cases follow a very are the object of study of iDEM since they will be assessed detailed prototyping process which will focus on colto implement our digital deliberative spaces. Our analy- lecting requirements from users and institutions, text sis so far has identified the following tools which could simplification functionalities, tools and formats such as benefit from the technology to be developed in iDEM: (i) questionnaires, for engaging hard to reach groups and Bang the Table 12 is a digital engagement platform that vulnerable populations, always considering a gender perenables communities to participate in decision-making. spective. Facilitators of the deliberative and participatory It provides tools for online discussions, surveys, and feed- processes will be interviewed in order to understand back that help decision-makers gather feedback and make the challenges of marginalized groups in processes of informed decisions. (ii) ChangeAView 13 is a platform political participation. The three use cases will be run that facilitates respectful, evidence-based discussions. It in sequence to allow feedback to be integrated into the encourages users to change their views on a topic after iDEM solution (technological and methodological) – See considering diferent perspectives, facts, and arguments. Figure 1. A thorough evaluation of the use cases by exter(iii) CitizenLab 14 is a platform that empowers citizens to nal experts will be carried out to asses the inclusiveness participate in local democracy. It enables users to propose and the accessibility of the proposed solution. ideas, vote on proposals, and engage in discussions with local oficials and community members. (iv) Decidim 15 is 7. Outlook an open-source participatory democracy platform that enables citizens to co-create policies and make collec- To be able to make informed decisions and actively get tive decisions. It provides tools for deliberation, voting, involved in society, people need to understand written inand proposal creation. (v) Consul 16 is an open-source formation, especially information to participate in demoplatform for citizen participation and engagement in gov- cratic processes which afect their lives. Unfortunately ernment. It provides tools for proposal creation, voting, information used by policy-makers and democratic inand deliberation, as well as open data and analytics for stitutions requite high literacy levels. Although several decision-makers. And (vi) Reflect! 17 is a platform for on- organizations ofer accessible information in many counline deliberation that enables participants to share their tries, they depend on well-trained human editors who opinions, discuss ideas, and propose solutions through can only produce a handful of documents at the time and moderated discussions. Some of the above tools are well at a high cost. The iDEM project will help improve the acestablished such as Decidim, some are open-source, and cessibility to information in democracy by creating novel only a few ofer accessibility functionalities (e.g, Citizen- participatory spaces in which text simplification technolLab). An example of the use of on-line platforms for ogy an natural language generation will be used to adapt deliberation and participation is the Conference of the texts or generate new ones thus facilitating participation Future of Europe, which, empowered by the European in democratic processes.
Decidim ecosystem, enabled citizens from all over the EU to share their ideas and points of view via online events.
However, in spite of its accessibility to information in 8. Lay Summary multiple languages of the EU, there was no support for engaging underrepresented groups such as people with reading and writing disabilities.
12https://engage.bangthetable.com/ 13https://www.bbc.com/news/blogs-trending-48579597 14https://www.citizenlab.co/ 15https://www.decidim.barcelona/ 16http://www.consulproject.org/en 17https://reflect.gatech.edu/
There are silenced voices in democratic spaces as millions of people struggle with language barriers, and are excluded from key deliberative processes. In the European Union (EU) alone, around 6 million individuals, and globally over 90 million, have dificulties to read, write, and comprehend language, which restricts their full participation in democracy. In this context, the EU-funded iDEM project aims to break these linguistic barriers and begin a new era of inclusive and participatory democratic spaces for marginalised communities. Specifically, the project will create next-generation multilingual models, automatically adapting texts to the needs of the people, simplification and its evaluation, in: ACL-IJCNLP and provide Artificial Intelligence tools for unbiased com- 2015, volume 2, 2015, pp. 823–828. munication. iDEM’s overall goal is to promote inclusivity [11] Y. Dong, Z. Li, M. Rezagholizadeh, J. C. K. Cheung, and representation for marginalised groups. EditNTS: An neural programmer-interpreter model for sentence simplification through explicit editing, in: Proceedings of the 57th Annual Meeting of the Acknowledgments Association for Computational Linguistics, Association for Computational Linguistics, Florence, Italy, This document is part of a project that has received fund- 2019, pp. 3393–3402. ing from the European Union´s Horizon Europe research [12] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, and innovation program under the Grant Agreement No. L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, At101132431 (iDEM Project). Views and opinions expressed tention is all you need, in: Advances in Neural are however those of the author(s) only and do neces- Information Processing Systems 30: Annual Consarily reflect those of the European Union. Neither the ference on Neural Information Processing Systems European Union nor the granting authority can be held 2017, December 4-9, 2017, Long Beach, CA, USA, responsible for them. UOL was funded by UK Research 2017, pp. 5998–6008. and Innovation (UKRI) under the UK government’s Hori- [13] F. Alva-Manchego, J. Bingel, G. Paetzold, C. Scarton, zon Europe funding guarantee (grant number 10103529). L. Specia, Learning how to simplify from explicit labeling of complex-simplified text pairs, in: ProceedReferences ings of the Eighth International Joint Conference on Natural Language Processing, Taipei, Taiwan, [1] E. Beauvais, Deliberation and equality, in: The Ox- 2017, pp. 295–305.
ford Handbook of Deliberative Democracy, Oxford [14] K. C. Sheang, H. Saggion, Controllable sentence simUniversity Press, 2018. plification with a unified text-to-text transfer trans[2] S. Gherghina, M. Mokre, S. Miscoiu, Deliberative former, in: Proceedings of the 14th International democracy, under-represented groups and inclu- Conference on Natural Language Generation, Assiveness in europe, Innovation: The European Jour- sociation for Computational Linguistics, Aberdeen, nal of Social Science Research 34 (2021) 1–5. Scotland, UK, 2021, pp. 341–352. [3] P. Orero, A. Matamala, Standardising accessibil- [15] C. Rafel, N. Shazeer, A. Roberts, K. Lee, S. Narang, ity: transferring knowledge to society., Journal of M. Matena, Y. Zhou, W. Li, P. J. Liu, Exploring the Audiovisual Translation 1 (2018) 139–154. Limits of Transfer Learning with a Unified Text[4] A. Matamala, Ó. García Muñoz, Easy Language in to-Text Transformer, Journal of Machine Learning
Spain, 2021, pp. 493–526. Research 21 (2020) 1–67. [5] H. Saggion, Automatic Text Simplification, vol- [16] S. Bott, L. Rello, B. Drndarevic, H. Saggion, Can ume 10 of Synthesis Lectures on Human Language spanish be simpler? LexSiS: Lexical simplification Technologies, Morgan & Claypool Publishers, 2017. for spanish, in: COLING 2012, 2012, pp. 357–374. [6] M. Shardlow, A Survey of Automated Text Simpli- [17] J. Qiang, Y. Li, Y. Zhu, Y. Yuan, X. Wu, LSBert: ifcation, International Journal of Advanced Com- Lexical Simplification Based on BERT, IEEE/ACM puter Science and Applications 4 (2014). Transactions on Audio, Speech, and Language Pro[7] García Muñoz, Léctura fácil - Métodos de redacción cessing (2020) 3064–3076.
y evaluación, Real Patronato sobre Discapacidad, [18] K. C. Sheang, H. Saggion, Multilingual controllable 2012. transformer-based lexical simplification, Proces. [8] H. Saggion, S. Štajner, S. Bott, S. Mille, L. Rello, del Leng. Natural 71 (2023) 109–123.
B. Drndarevic, Making It Simplext: Implementa- [19] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. tion and Evaluation of a Text Simplification System Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, for Spanish, ACM Transactions on Accessible Com- G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, puting 6 (2015) 1–36. G. Krueger, T. Henighan, R. Child, A. Ramesh, [9] W. Coster, D. Kauchak, Learning to simplify sen- D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, tences using Wikipedia, in: Proceedings of the E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, Workshop on Monolingual Text-To-Text Genera- C. Berner, S. McCandlish, A. Radford, I. Sutskever, tion, Association for Computational Linguistics, D. Amodei, Language models are few-shot learners, Portland, Oregon, 2011, pp. 1–9. URL: https://ac in: H. Larochelle, M. Ranzato, R. Hadsell, M. Ballanthology.org/W11-1601. can, H. Lin (Eds.), Advances in Neural Information [10] S. Štajner, H. Béchara, H. Saggion, A deeper ex- Processing Systems, volume 33, Curran Associates, ploration of the standard PB-SMT approach to text Inc., 2020, pp. 1877–1901.