Democracy made easy: simplifying complex topics to enable democratic participation

Proceedings of the 1st Workshop on Artificial Intelligence and Easy and Plain Language in Institutional Contexts (AI & EL/PL), pages 108–124, June 23, 2025 Democracy Made Easy: Simplifying Complex Topics to Enable Democratic Participation Nouran Khallaf1, Stefan Bott2, Carlo Eugeni1, John O’Flaherty3, Serge Sharoff1,Horacio Saggion2 1School of Languages, Cultures and Societies, University of Leeds, United Kingdom 2Department of Engineering, Universitat Pompeu Fabra, Spain 3The National Microelectronics Applications Centre (MAC) Ltd, Ireland Abstract Several groups of people are excluded from democratic deliberation because the language used in this context may be too difficult for them to understand. Our iDEM project aims to reduce existing linguistic barriers in deliberative processes by developing technology to facilitate the translation of complicated text into Easy-to-Read formats that are more suitable for many people. In this paper, we describe classification experiments for detecting different types of difficulties which should be amended in order to make texts easier to understand. We focus on a lexical simplification system that can achieve state-of-the-art results with the use of a free and open-weight large language model for the Romance Languages in our project. Moreover, a sentence segmentation system is introduced to produce text segmentation for long sentences based on training data. Finally, we describe the iDEM mobile app, which will make our technology available as a service for end-users of our target populations. 1 Introduction Representative democracy is based on delegating policy matters to elected representatives, while the deliberative democratic process aims at involving the stakeholders directly (Bächtiger et al.,2018). Modern democratic institutions aim at a greater focus on stakeholders’ involvement. However, this has the requirement of clearer language, which is accessible to the stakeholders, especially in cases where the stakeholders face challenges in understanding, for example, in such cases as people with intellectual disabilities or non-native speakers. The demand for better communication is also reflected in the international treaties, in particular, the Universal Declaration of Human Rights, in its Article © 2025 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CCBY-ND. 19, affirms everyone’s right to seek and receive information. Moreover, particularly important in this context is the United Nations Convention on the Rights of Persons with Disabilities (CRPD), which includes accessibility as one key enabler for a more inclusive society. That is, the ability of any product, service, content, environment, etc., to be used by people with the widest range of abilities (including linguistic and cognitive abilities). The CRPD also considers accessibility as, for example, an enabler for democratic participation rights such as freedom of expression and self-determination. Consequently, a lack of accessibility can be linked to a risk of exclusion for persons who cannot participate equally due to linguistic barriers. The focus of our paper is on providing an introduction into technologies developed in the context of our project, iDEM 1 2 : in the area of intersectionality and equality in deliberative and participatory democratic spaces, iDEM aims at making information more accessible and inclusive in the context of democracy and in particular in deliberative and participatory processes. More specifically, in this paper, we will discuss: 1. A tool for assessing sentence-level complexity and predicting appropriate simplification strategies; 2. Applications of these tools to real-world corpora, including the United Nations Parallel Corpus and Europarl; 3. A text simplification pipeline powered by Large Language Models (LLMs), focusing on lexical simplification and Easy-to-Read (E2R) sentence segmentation. The rest of the paper is structured as follows: 1 Innovative and Inclusive Democratic Spaces for Deliberation and Participation 2https://idemproject.eu/ 108 Section 2provides an overview of the iDEM project. Section 3reviews related work on complexity assessment and text simplification. Section 4 details our sentence complexity classifier and simplification approach, including the evaluation results. Section 5outlines the mobile application in development, while Section 6discusses current limitations. Finally, Section 7offers concluding remarks. 2 Project Overview The iDEM Project in the area of intersectionality and equality in deliberative and participatory democratic spaces aims at making information more accessible and inclusive in the context of democracy and, in particular, in deliberative and participatory processes. In the first phase of the project we have investigated, using a theoretical approach, current marginalization from deliberative processes of diverse underrepresented groups due to language skills in order to understand what are the linguistic barriers which hamper their participation. By working with different associations, iDEM adopts a user-centered approach in use case design and data collection and annotation to ensure maximum impact in the community, thus contributing to making democracy more accessible and inclusive. An innovative iDEM service (e.g., mobile app) is being implemented to host the developed language technologies to support on-demand simplification in Catalan, English, Italian, and Spanish. In the current phase of the project, we are developing the underlying natural language processing technology as well as fine-tuning the use cases to test and evaluate our proposed approach to a more inclusive deliberative democracy. The interested reader is referred to (Saggion et al.,2024b) for an overview of the project. 3 Related Work 3.1 Easy-to-Read Since the late nineties, many organisations have raised awareness about fundamental information being written in a way that is too difficult to understand for many people. Initiatives to palliate this deficit include “Plain Language” (U.S. Government,2011) and “Easy-to-read” (Inclusion Europe, 2009). They are two different methods whose overall objective is to make information more accessible. They proposed guidelines for how to write more accessible texts; however, applying them to produce accessible material heavily depends on well-trained human editors and, therefore, considerably limits the production of easy-to-read or plain language texts. Compared to standard language, easy-to-read language is a simplified version for the sake of readability for specific audiences (Caro,2017). In this paper, we adopt E2R over Plain Language because its structured guidelines form the foundation of diverse and adaptable translation strategies designed to make information accessible to people with reading difficulties, including people with intellectual disabilities. They have little command of the language and poor literacy. Empirical research in the field is uncommon (González-Sordé and Matamala,2024), especially when compared to fields such as automatic text simplification. Although the topic has gained greater scholars’ attention in recent years, sometimes research reports on apparently contradictory findings (Fajardo et al.,2014) between guidelines and actual text understanding by target E2R populations; moreover, even guidelines appear to take on different aspects with little agreement between them (Maaß,2020). With the advances that natural language processing has achieved in recent years, interest in the automatic adaptation of texts to plain language or E2R has intensified (Alarcon et al., 2021;Da Cunha Fanego,2021;Saggion,2024). 3.2 Complexity Identification The first focus of our research within iDEM is on theoretically understanding the factors that contribute to the complexity of a text or the sentences within this text. The guidelines described in the previous section are directed to human editors and often leave much room for interpretation and are hard to operationalise, for example the instruction to avoid difficult words. We are interested in combining theoretical insights with data-driven analysis and classification. In previous work, computational studies typically overlook insights from translation studies, particularly the various strategies proposed (Vinay and Darbelnet,1971;Newmark,1988;Chesterman, 1997;Zabalbeascoa,2000;Molina and Hurtado Albir,2002;Gambier,2006), focusing on the systematic processes involved in translating a source text into a target text across languages. Translation studies provide a complementary approach in examining strategies used in intra-lingual translation, where a source text is translated into a target text 109 in the same language. Eugeni and Gambier (2023) argue that such transfers habitually achieve a complete correspondence between source and target texts. One key task in order to transform sentences into E2R is lexical simplification, i.e., simplifying individual words or short phrases independent of the effect of such simplifications on the overall sentence coherence. For instance, Paetzold and Specia (2016) developed methods that specifically targeted complex word identification (CWI), which detects difficult words and suggests simpler alternatives. These techniques usually ignore how such simpler words would fit the general sentence structure. Datasets developed to evaluate lexical simplification, e.g., SemEval-2012 Task 1 (Specia et al., 2012), ALEXSIS (Ferrés and Saggion,2022), TSAR 2022 (Saggion et al.,2022) or MLSP 2024 (Shardlow et al.,2024) have aided a focus on single word-level replacements. Though helpful, these datasets primarily cover single word substitutions in isolation rather than more general context-sensitive simplifications. As a result, simplifications generated with the assistance of these tools sometimes sound unnatural, which needed a post generative model to refine sentence coherence. This issue was also highlighted by Shardlow (2014), who reviewed various lexical simplification approaches and noted that, while effective for readability, they frequently ignore sentence coherence and grammatical correctness. Corpora for sentence simplification includes ASSET (Alva-Manchego et al.,2020) that provides multiple quality simplifications per sentence. However, ASSET still focuses to some extent on fine-grained lexical or phrase-level modifications and lacks annotations for deeper grammatical or discourse-level modifications. Similarly, WikiLarge (Zhang and Lapata,2017) provides large parallel sentence pairs for training simplification models but does not explicitly annotate the simplification strategies, making it difficult to study in detail exactly how sentences are simplified. The Simplext corpus (Saggion et al.,2015) provides full document simplifications following E2R guidelines for the Spanish language without indication of transformation type while PorSimples (Aluísio and Gasperin,2010) provides document simplification in Portuguese covering several operations. 3.3 Text Simplification Our focus for this paper is on lexical simplification; for an overview of full text simplification approaches and methods, the reader is referred to (Saggion,2017). Several past approaches to lexical simplification used traditional count-based word-vectors and available dictionaries for modelling word semantics and to select simple word replacements for complex words (Biran et al.,2011; Bott et al.,2012); in later work, word embedding were used, which is learned from huge text collection (Glavaš and Štajner,2015). More recently, large-scale language models such as BERT and its variations have been applied to predict substitution candidates for complex words. For example, LS-BERT (Qiang et al.,2020) uses the masked language model (MLM) of BERT to predict a set of candidate substitution words and their associated probability. In this context, the MLM predicts substitute words which are ranked for simplicity using: probabilities, a language model, a paraphrase database, word frequency and word semantic similarity with the target word. Very recent work presented in the TSAR 2022 (Saggion et al., 2022) and MLSP 2024 (Shardlow et al.,2024) evaluation frameworks have demonstrated that Large Language Models (LLMs) are in fact the best performing models for the lexical simplification. Techniques such as “prompting” are used to condition the LLMs to produce a simplification or to suggest alternative words. Note, however, that these models underperform when simplifying low-resourced languages. We define ’low-resourced languages’ as those with limited digital text corpora (e.g., Catalan vs. English), impacting LLM performance as noted in Section 4.4. Despite advances in lexical simplification (e.g., TSAR 2022, MLSP 2024), key gaps remain: (1) How can simplification strategies be systematically categorised beyond lexical substitution? (2) What taxonomies exist for intra lingual translation, and how do they apply to automation? Section 4.2 addresses these by proposing a strategy taxonomy, testing it on institutional corpora, and leveraging LLMs without prompt engineering—a less-explored approach due to its complexity (Shardlow et al.,2024). 4 Natural Language Processing for Easy-to-Read Translation 4.1 Datasets We use a range of datasets across different components of our system 3 . The primary dataset used 3 Where applicable, datasets are available on request from the authors or are publicly accessible through the cited sources. 110 for complexity assessment and simplification strategy classification comprises 76 parallel texts collected from Scottish care services, UK political manifestos (2024), and Disability Equality Scotland newsletters. These cover diverse topics such as healthcare, environmental policies, disability advocacy, and accessibility. The texts were manually aligned at the sentence level, resulting in 4,166 words in 206 original (“complex”) sentences and 3,259 words in 210 simplified counterparts. Despite the reduction in word count, the number of sentences increased slightly, reflecting a key simplification strategy that is splitting longer sentences to improve readability. We additionally use a French dataset of 370 manually aligned sentence pairs. The original texts were retrieved from the Réfugiés.info website and were anonymised to remove any personally identifiable information (PII) (Team,2025). These parallel sentence pairs provide training data for our simplification strategy classifier (Section 4.2).4 For evaluating our system on larger, multilingual corpora, we use the European Parliament (Koehn,2005) and the United Nations Parallel Corpus (Ziemski et al.,2016). These are publicly available and provide high-quality sentencealigned translations in English, Spanish, and Italian. We applied our multilingual classifier to these datasets to analyse simplification needs across languages (Section 4.3). For lexical simplification, we use few-shot prompting on pre-trained Salamandra models with trial data from the MLSP 2024 shared task (Shardlow et al.,2024), covering English, Spanish, Italian, and Catalan (Section 4.4). Finally, for sentence segmentation in Spanish according to E2R standards, we accessed a private annotated dataset provided by Calleja et al. (2024). This dataset includes 3,826 training, 484 validation, and 1,452 development sentences, each annotated with E2R-compatible cut points (Section 4.6). 4.2 Complexity Assessment The simplification strategy prediction task aims to determine the types of transformations needed to make a sentence more accessible. Table 1provides examples of these transformations. 4 English dataset was annotated by a linguist with expertise in translation and text simplification, using the same predefined set of simplification strategy categories described in Appendix B; the French dataset was labelled by the Réfugiés.info editorial team following the same guidelines and category definitions. Our taxonomy is informed by Inclusion Europe’s guidelines (Inclusion Europe,2009), intralingual translation practices into E2R (Hansen-Schirra et al.,2020), and a qualitative analysis of our dataset. While previous taxonomies in Translation Studies have offered valuable models for interlingual and diamesic translation, they lack the granularity needed to describe all strategies observed in E2R practice. On the other hand, typologies in Automatic Text Simplification (ATS) are based on corpus analysis (Bott and Saggion,2014) or on edit operations that mainly deal with adding, deleting, replacing, and moving words Cardon et al. (2022). However, texts translated in E2R language clearly show that professionals in the field apply many more operations that pertain to the field of pragmatics and semiotics, focused on how concepts are distributed and or explained to help the user understand them. To address this gap, our framework adapts insights from both domains. Based on Inclusion Europe’s three levels of simplification—lexical, syntactic, and semantic—we define eight macrostrategies that range along a continuum from additive operations (e.g., Explanation) to reductive ones (e.g., Omission). These are outlined in Table 2 comprises 8 macro-strategies (excluding transcript since it is a non-simplification operation), 8 strategies, and 30 micro-strategies. For the full set of strategies, see Table 10 (Appendix E). Cross-linguistic differences in simplification strategies are also relevant. In our multilingual experiments, we observed variations in dominant strategies across English, Spanish, and Italian, which suggests that language-specific features influence how simplification is operationalised. This will be further explored in Section 4.3. The classifier is built by application of pretrained transformer-based models (such as multilingual BERT (Devlin et al.,2019)) for multiclass text classification, focusing on the prediction of simplification strategies need to simplify the respective sentences. We employed Stratified 5-fold Cross-Validation for rigorous evaluation and generalisation. We took the average of the validation scores across all the folds to determine the final scores. Early stopping was also employed, wherein the training was halted if the validation loss did not see an improvement for the patience period. Class imbalance in the data, with certain strategies being underrepresented, was a problem during training. To counter this, we used a weighted 111 In 2018-20 life expectancy at birth in Scotland was 76.8 years for males and 81.0 years for females. From 2018 to 2020 babies born in Scotland were expected to live 77 years if they were boys and 81 years if they were girls. Modulation Explanation Synonymy,Syntax Synonymy,Syntax Table 1: Segment alignment for the original (top) and simplified (bottom) sentences Strategy Description Example Omission Removing unnecessary rhetorical or diamesic constructs. “Sir Keir Rodney Starmer KCB KC” →“Starmer” Compression Simplifying grammatical/semantic constructs. “to guide the group” →“to the group” Syntactic Change Adjustments between syntactic levels. “citizens” →“people in Scotland” Transcript No changes made to the text. “I love music” Transposition Word class change. “our aim is” →“we want” Synonymy Simplifying technical or abstract words. “conversation” →“talk” Modulation Redistributing information linearly. “joins in activities... supported by family” → “He joins activities. His family helps.” Explanation Making hidden content or terms explicit. “co-design services...” → “co-design means sharing your ideas” Illocutionary Change Making implied meaning explicit. “know your body’s library” →“know your body” Table 2: Simplification strategies required for a sentence, with examples cross-entropy loss function. Class weights were calculated as the inverse frequency of each class: wc= 1 freqc ·N 2(1) where wc is the weight assigned to class c , freqc is the frequency of class c , and N is the number of samples. This approach ensured that underrepresented classes contributed more to the overall loss, so the model became more capable of predicting the minority classes. Additionally, gradient clipping was applied during training to stabilise the optimisation. Gradient clipping limits the maximum value of gradients during backpropagation to prevent extremely large updates of model parameters that could destabilise training or lead to divergence. Mathematically, gradient clipping can be expressed as: gclipped = min g, gthreshold ∥g∥,(2) where g represents the original gradient vector, gthreshold is the clipping threshold, and ∥g∥ is the norm of the gradient vector. Gradient clipping ensures consistent updates to model parameters, improving training stability. See the summary of hyper-parameters in Table 8 (Appendix B). The use of medium-sized PLMs (such as multilingual BERT) instead of LLMs helps with the possibility of applying the classifiers to large institutional datasets (such as the entirety of Europarl or the United Nations Corpus), as well as with the possibility of deploying the classifiers to guide the corrections. We used standard precision, recall, and F1-score metrics (Manning et al.,2008) to evaluate model performance. Given the class imbalance, we report the weighted macro F1-score (Sokolova and Lapalme,2009), which better reflects the classifier’s ability to handle both frequent and rare simplification strategies. The fine-tuned classifier model achieved a weighted macro F1-score of 0.8089, demonstrating its ability to generalize across majority and minority classes. In particular, it outperformed the baseline majority-class strategy, which corresponds to the weighted macro F1-score of 0.096. The F1 score of the multilingual model (trained on English, tested on French) is 0.6339, thus reflecting the need to improve its ability to generalize across languages. However, given that its errors are balanced, i.e., the model is confused with predicting Synonymy for Explanation and vice versa, see the confusion matrix in Figure 2(Appendix C). Omission and Compression categories tend to be confused with one another, with Omission commonly predicted as Explanation or Transcript, mirroring the need to enhance the separation between removal and rewriting strategies. Modulation is also commonly confused with Synonymy, mirroring the need to strengthen sentence restructuring cues in training. 112 Category English Spanish Italian # Sent. % # Sent. % # Sent. % Europarl Total Sentences 2,005,688 100 1,788,913 100 1,928,874 100 Complex 1,932,492 96.3 1,660,631 92.8 1,868,714 96.8 Omission 59,065 3.1 23 0.001 57 0.003 Syntactic Change 254,483 13.2 11,777 0.7 21,321 1.1 Transposition 13,075 0.7 35,053 2.1 40,633 2.2 Synonymy 1,104,564 57.2 37,259 2.2 81,468 4.4 Modulation 41,802 2.2 724,469 43.6 1,004,438 54.2 Explanation 459,503 23.8 852,050 51.3 702,526 37.9 UN Corpus Total Sentences 10,600,000 100 10,665,709 100 Complex 9,628,533 96.2 9,987,750 93.6 Omission 75,217 0.7 62 0.0006 Syntactic Change 181,228 1.8 503,047 5.0 Transposition 39,356 0.4 68,878 0.7 Synonymy 4,587,340 45.0 198,479 1.9 Modulation 445,095 4.3 5,345,515 53.5 Explanation 4,878,679 47.7 3,871,769 38.7 Table 3: Sentence counts and proportions of simplification strategies in institutional datasets 4.3 Experiments with assessing institutional repositories We experimented with two institutional repositories, which include English, Italian and Spanish, some of the languages of our project, the corpus of the European Parliament (Koehn,2005) and the United Nations Parallel Corpus (Ziemski et al., 2016). Both resources include high-quality translations, so the content of each sentence is the same. However, we can expect that the three languages differ in their traditions for maintaining linguistic complexity in such formal texts as the parliamentary proceedings. Total sentences row in Table 3 presents the amount of data in each dataset. We used sentence-aligned versions from the respective repositories and applied the multilingual classifiers described in the previous section to make predictions. If the complexity classifier detected the need to simplify the sentence, i.e., it was predicted as "Complex", we estimated the likely strategy needed for this task. As the classification model is limited to the one-label setup, out of several edit operations required for a sentence (see the example in Table 1), our current version of the model predicts the single most likely operation (Explanation in this example). Table 3shows that the majority of sentences in both datasets and across all the languages considered (English, Spanish, and Italian) require some form of simplification. For English sentences, the most common simplification operations found are (1) lexical substitution (synonymy), primarily through the choice of simpler synonyms, and (2) Explanation which provides more explanation to facilitate reading. Conversely, for both datasets of Spanish and Italian sentences, the predominant simplification strategy is modulation, with a particular emphasis on sentence restructuring for the purpose of achieving a more linear and straightforward reading experience. 4.4 Simplifying Complex Words As reported in recent lexical simplification challenges (i.e. TSAR 2022 (Saggion et al.,2022) and MLSP 2024 (Shardlow et al.,2024)), most recent state-of-the-art lexical simplification systems rely on decoder-only autoregressive LLMs like GPT-4 (Enomoto et al.,2024). These systems seem to systematically outperform other systems, like encoderonly language models (e.g. BERT), also because recent developments of LLMs have mostly concentrated on decoder models. Decoders are generally more flexible and have strong zero-shot or few-shot abilities. Commercial closed-weight models like GPT-4, however, carry concerns for the purpose of our project since they lack guarantees of privacy protection and generate costs by using the API. In addition, their closed nature does not usually allow us to fine-tune them. In preliminary experiments, we found out that the LLMs of the Salamandra family 5 (GonzalezAgirre et al.,2025) perform very well on European Languages, especially on Romance languages, and within the last group, they especially excel at the performance of Catalan. This can be explained because Salamandra models are part of the Alia initiative (Government of Spain) funded by the Spanish government with a strong focus on languages spoken in Spain. Salamandra models were trained as decoder-only, and they are also provided as instruction-tuned versions. With this, we decided to use a simple few-shot system as our first approach. Few-shot prediction from a pre-trained model refers to the process where a model that has already been trained on a large dataset (a pre-trained model) is used to make predictions or perform tasks with no or only a few labeled examples for a new task. The shots are examples provided in the prompt, as opposed to being used as training data for fine-tuning. Zero-shot prediction does not provide any example. The pre-trained model is typically a decoder-only model, which produces output based on an input prompt that conditions 5https://huggingface.co/collections/BSC-LT/ salamandra-66fc171485944df79469043a 113 the output. In essence, few-shot prediction from a pretrained model means leveraging a model’s prior knowledge from a large dataset to perform well on a new task or dataset, even with very few labeled examples. As our pre-trained models, we used the 2 billion parameters (2B) and the 7 billion parameter (7B) versions of Salamandra. We used the following prompt without doing refinement through prompt engineering: Given the context and the specified target word in {LANGUAGE}, answer 10 simpler alternative words. Do not give less than 10 alternative words. Give different words as alternatives. {SHOT_EXAMPLES} Context: {CONTEXT} Target Word: {TARGET} Alternatives Words: Here LANGUAGE is a variable which is set according to the language (Catalan, English, Italian, Spanish) in which we want to produce predicted solutions. For few-shot prediction we used examples from the trial section of the MLSP data. The shot examples were selected randomly, but we made sure that unique contexts were selected. An instance of a SHOT_EXAMPLE is given here: ´ Given the context.... Context: A continue statement will skip the remainder of the block and start at the controlling conditional statement again. Target Word: remainder Alternative Words: rest, restrictive, remaining, remainder, balance For a 2-shot or 4-shot prompt, 2 or 4 of these different examples would be included in the prompt given to the system. The CONTEXT and TARGET variables have the same form as in the provided shot examples. As evaluation measures, we used the same as in the MLSP shared task (see Section 3.3). Accuracy (ACC) expresses the percentage of right solutions given out of all given solutions. Here we use Accuracy@1@top1 which is defined as the percentage of instances where the first top-ranked substitute matches the most frequently suggested synonym in the gold data (top1). MAP@k (Mean Average Precision) uses a ranked list of generated substitutes, which can either be matched or not matched against the set of the gold-standard substitutes. The first ksolutions of the ranked list are considered. The results can be seen in Table 4. We use the same baseline here as was used in the MLSP shared task. It has to be noted that the baseline used there was very strong, since it used zero-shot prompting with the use of the chat-fine-tuned version of Llama-2-70B. This is a version with 70 Billion Parameters and thus ten times larger than the Salamandra-7B model we use here. In fact, many participating systems in the MLSP shared task could not outperform this baseline. In the tables, we mark those results with an asterisk that are higher than this baseline. As a further reference we also list the performance of the different winning systems of the shared task. These winners, however, use GPT-3 for Catalan (Dutilleul et al.,2024) and GPT-4 (Enomoto et al.,2024) for the rest of the languages, and for reasons we describe above, we cannot use them for the iDEM project. As expected, the 2B version of Salamandra could not outperform the baseline (Table 9in Appendix D). We attribute this to the fact that this model is too small to produce reliable results in a task that requires quite a large amount of general knowledge about language, such as synonymy and simplicity. The results from this table are still interesting because we want to use fine-tuning on Salamandra2B in future work. The 7B version of Salamandra, on the other hand, could outperform the baseline nearly systematically in few-shot settings. Interestingly, the difference between 2-shot and 4-shot predictions is not very large. In some cases, the 4-shot predictions perform even worse than 2-shot predictions. Another observation that can be made is that Salamandra mostly excels at the three Romance languages Spanish, Catalan and Italian, while for English, it performs very close to the baseline. In this case, it means that the baseline is higher and harder to beat for English than for the other languages because of the multilingual capabilities of the baseline system or the lack thereof. These observations confirm our assumption that Salamandra is a good choice for the set of languages that we have to treat in iDEM. 4.5 Integration of Complexity Assessment and Lexical Simplification This section presents ongoing work toward integrating two core modules of our system: complexity assessment (Section 4.2) and lexical simplification (Section 4.4). The classifier first detects complex lexical items in a sentence, and the simplification module then proposes easier alternatives. While the full pipeline has not yet been formally evaluated, we have implemented a proof-of-concept 114 0-Shot 2-Shot 4-Shot MLSP Baseline MLSP Winner ACC MAP@3 ACC MAP@3 ACC MAP@3 ACC MAP@3 ACC MAP@3 English 0.1280 0.1912 0.4017* 0.3868 0.4035* 0.4242* 0.3877 0.4241 0.5245 0.5762 Spanish 0.0286 0.1213 0.3541* 0.5148* 0.3608* 0.3644 0.3254 0.4157 0.4536 0.6763 Catalan 0.0426 0.1390 0.2292* 0.3742* 0.2022* 0.3357* 0.1977 0.3024 0.2719 0.5003 Italian 0.035 0.1419 0.3596* 0.4108* 0.3315* 0.3868* 0.2964 0.3310 0.4762 0.5661 Table 4: Results of Zero and Few Shot Lexical Simplification Performance for a big model (Salamandra-7B). Results are compared to the state of the art as reported in the recent MLSP 2024 lexical simplification shared task. Asterisks (*) indicate the model outperformed the strong baseline of the competition. Sentence Easy to Read Segmentation The way this sentence is cut is easy to read. The way this sentence is cut is easy to read. Validar es comprobar si un documento es fácil de comprender. Validar es comprobar si un documento es fácil de comprender. Table 5: Examples of segmented sentences in English and Spanish taken from Easy-to-Read guidelines. to illustrate its feasibility. Table 6provides multilingual examples where the complexity classifier flags difficult words, which are then simplified by the Salamandra-7B lexical simplifier. For instance, in the English sentence “The reason why hypothalamic lesions affect body fat...,” the words ‘hypothalamic’ and ‘lesions’ are identified as complex and replaced with ‘brain’ and ‘damage,’ respectively—substitutions that significantly enhance readability. In the context of the iDEM project, this integration is intended for deployment within the mobile application currently under development (see Section 5), where users with cognitive or linguistic barriers can receive real-time support in understanding complex information. Future work will involve formal evaluation, expansion to full sentences, and deeper cross-linguistic adaptation. 4.6 Segmenting Sentences for Easy-to-Read According to E2R standards (Inclusion Europe, 2009;153101,2018), sentences in E2R are recommended to be short and should fit on one line on the printed page (or screen). Since this is not always possible, guidelines recommend cutting the sentence where people would pause when reading out loud. Research on sentence segmentation is somehow related to the prediction of prosodic markers in text-to-speech systems, where syntactic structure and word/token information is used (Fitzpatrick and Bachenko,1989). Examples of how sentences should be segmented in E2R in English and Spanish are presented in Table 5. Although datasets for sentence and lexical simplification exist (as reported above), there is a lack of publicly available datasets of E2R segmentation. We have gained private access to a dataset of segmented E2R texts in Spanish (Calleja et al., 2024). This dataset is organized into three files corresponding to train (3,826 sentences), validation (484 sentences), and development (1,452 sentences). Each sentence is explicitly marked to indicate where it should be segmented following E2R standards. We adopt a machine learning approach to sentence segmentation, developing a classifier based on linguistic information and other features such as the position of the token in the sequence (first, second, etc.) or the distance to the previous cut. We process the dataset in order to convert the original sentences into instances for learning. The instances for learning are based on the tokens (words, punctuation, numbers, etc.) in each sentence; our aim is to classify all tokens as cut-point or not. In order to create the learning instances, we linguistically analyze each sentence using a Spanish model from the SpaCy library (Honnibal et al.,2020), which produces information on parts of speech, syntactic dependencies, and named entities. We extract several features including the Parts Of Speech (POS) tag of the token, the case of the token (lower cased, upper cased), whether the token is a punctuation, whether the token is part or a named entity (begin, inside, outside), the position of the token in the sentence, the distance to the previous cut point (or -1), and the distance to the end of the sentence. The learning instances (one per token) are stored in a CSV file for use by a machine learning algorithm. We report results using a Decision Tree algorithm (Steinberg,2009) due to its simplicity and explanatory power (i.e. set of rules). Other algorithms were less successful on our data. The learning algorithm is an instance from the De115 Lang Context (Sentence) Complex word (by CA) Substitute (by LS) Eng The reason why hypothalamic lesions affect body fat and feeding behavior has in fact much to do with leptin signaling. hypothalamic brain lesions damage affect influence Spa Si este indicador baja de 1, implicaría que la empresa no está en capacidad de cubrir sus obligaciones de corto plazo con los activos líquidos que posee. (If this indicator is below 1, this implies that the enterprise is not in conditions to cover its obligations in the long run with the liquid assets it possesses.) implicaría significaría indicador medida plazo tiempo activos bienes Cat La formació sosté que "els posicionaments excloents en vers a altres realitats educatives fonamentades amb idees polítiques distorsionen la realitat del model" català. (The formation maintains that "exclusionary positions towards other educational realities based on political ideas distort the reality of the Catalan model".) sosté defensa posicionaments posicións vers contra Table 6: Examples of cases where the Complexity Assessment (CA) system identifies a word that needs simplification and the Lexical Simplification (LS) system simplifies it. cision Tree implementation provided by the Scikit Learn library6(Pedregosa et al.,2011). Table 7reports segmentation results for the decision tree classifier and two baselines. The baselines are based on (i) the Parts of Speech (POS) tag, which on training data is the best predictor of the token where the sentence should be segmented, and (ii) on the most common length of the segment. As for the decision tree, two methods are applied: the oracle configuration knows about the true previous cuts, while the blind configuration has only access to the predicted previous cuts. The difference between oracle and blind configurations are expected. The difference in performance between the decision tree and the baselines is an indication that the features are contributing to the classification performance. Future work should look at analyzing feature contribution and improving the models, and providing segmentation support for Catalan, English and Italian. Algorithm F1 (Cut) F1 (No Cut) Avg. F1 Decision Tree (Oracle) 0.43 0.89 0.66 Decision Tree (Blind) 0.26 0.91 0.58 POS Tag Baseline 0.17 0.95 0.56 Seg. Length Baseline 0.12 0.91 0.52 Table 7: Segmentation results (based on F1 measure) into Easy to Read (Spanish data). Comparison of a Decision Tree with baselines. 5 Accessing Simplification Technology through the iDEM App The iDEM project implements and deploys a cloudbased, open-API iDEM platform to deliver text6https://scikit-learn.org/stable/modules/tree. html simplification services, integrating components for complex language detection (Section 4.2) phenomena and adaptation through text simplification (Section 4.4). It supports diverse audiences, languages, and domains, and solutions are made available for deliberative participatory spaces as open-source products. The current version of the app supports iteration via typed text, speech, OCR or PDF. A participation functionality allows the user to check proposals currently being discussed and simplify them for better understanding. For example, the Decidim platform (Aragón et al.,2017) can be directly accessed from the app to translate, or simplify active participatory processes. Examples of the APP in action can be seen in Figures 1a and Figures 1b. Note that the current simplification technology supported by the app is not yet the one described in the paper; it still serves as a demonstrator of what it will look like in the coming months. 6 Limitations and Ethical Considerations The studies on Complexity Assessment in section 4.2 and 4.3 argue for an analysis and simplification of a large array of factors, one of which is lexical simplification. We are aware that this is a current limitation, but future versions of the iDEM simplification tools will include a full treatment of sentence simplification. Our current simplification model, although achieving good performance in comparison with a strong baseline, does not do so with respect to the state of the art. This can be explained by our aim to keep models open and accessible to a broader community of stakeholders, i.e. lighter, open models could be afforded by more disadvantaged communities in the spirit of our project. Since our project deals with pro116 (continued from previous page) Strategy MacroStrategy Explanation and Examples Total ModLin Modulation Redistribution of sentence components: -ModWord: “...collaboration and information sharing...” → “...working together and sharing information...” -ModGrou: “Accessible Museums is a topic...” → “Our members think it is important to talk about Accessible Museums” -ModClau: “To improve community health... the Government works...” → “The Government works... to improve...” 2 PraSyn Synonymy Pragmatic synonyms: -PraProp: UN → United Nations, Nutella → chocolate cream -PraCont: “Sir Keir Starmer” → “the new Prime Minister” SemSyn Synonymy Semantic synonyms: -SemStere: ponder →think -SemHype: lecturers →teachers -SemHypo: flora →trees and flowers 3 GraSyn Synonymy Grammatical synonyms: -GraPron: “you don’t see it” → “you don’t see the mistake” -GraTens: “we have been doing” → “we have done” -GraPass: passive →active -GraNega: “not an obstacle” →“facilitated” TraNou Transposition Noun transposition. e.g. “our aim” →“we want” TraVer Transposition Verb transposition. e.g. “listening to music” →“music” TrAdje Transposition Adjective transposition. e.g. “mountainous landscapes” →“mountains” 4 TrAdve Transposition Adverb transposition. e.g. “behaving happily” →“was happy” Transcript Transcript A sentence is left unchanged. SynW2G/S/C Syntactic Change Word to group/clause/sentence SynG2W/C/S Syntactic Change Group to word/clause/sentence 12 SynC2W/G/S Syntactic Change Clause to word/group/sentence SynS2W/G/C Syntactic Change Sentence to word/group/clause Illocutionary Change Illocutionary Change Making implied meaning explicit. 1 GraSim Compression Grammatical simplification. e.g. “so as to” →“to” SemSim Compression Semantic simplification. e.g. condensing explanations 2 OmiEle Omission Omission of elements: -OmiSubj: “Sir Keir Rodney Starmer...” → “Starmer is...” -OmiVerb,OmiComp,OmiClau,OmiSent (e.g. full sentence removed) (continued on next page) 123 (continued from previous page) Strategy MacroStrategy Explanation and Examples Total OmiDia Omission Omission of discourse elements: -OmiFil: “you know” →removed -OmiRef: “I was tight... right when...” → “I was right when...” -OmiRhe: “wasn’t I?” →removed 2 Total 30 Table 10: Macro-strategies, Strategies, Micro-strategies, and Examples with Annotated Totals 124