The presence of unknown words in a text can have a significant impact on the reader's understanding, leading them to misinterpret the content. This challenge is magnified when dealing with complex words, whose meaning often depends on context and is not easy to infer. This problem extends to users who access texts issued by public institutions. In response to this need, we present GovAIEasy, a web application based on Artificial Intelligence, designed to simplify the understanding of institutional texts by applying GPT-4o and Zero-Shot learning. This application transforms complex texts into accessible and easy-to-understand versions for the user. In addition, the GPT-4o model is used to generate a definition, an example, and a use case of the detected complex word, facilitating understanding for users with a low literacy level, cognitive dificulties, or disabilities, helping them better understand the instructions and procedures required in public ofices.
lead to misinterpretations, disinterest, or even abandonment of reading [
Many people face significant reading comprehension barriers when dealing with public administration
text [
given target population is commonly known as complex word identification (CWI) [
CEUR Workshop
ISSN1613-0073
sionals and have opened a wide range of beneficial opportunities for its application in the public sector
[
The relevance of GovAIEasy1 lies in its valuable contribution to the assessment of lexical complexity in texts. Its main objective is to promote digital modernization, supporting the creation of more agile, open, and innovative governments. This tool ofers various applications relevant to research in the field of natural language processing and language simplification in government documents: • Automatic Complex Word Identification . Automates the process of identifying complex words in a text, facilitating the analysis and review of large volumes of content, which is especially useful in government documents. • Complexity Level Assignment. The system assigns a dificulty score to each complex word in the range of 0 to 1, classifying them into three levels: moderately dificult, dificult, and very dificult. This classification is essential to assess the impact of words on text comprehension. • Summary Generation. In addition, the model generates simple summaries of the Spanish text, facilitating a quick and accessible understanding of the content without having to read the full text. • Continuous improvement through data logging. Data on complex words and their classiifcation are stored in a database, allowing continuous monitoring of model performance and improving the accuracy of word identification and classification over time.
In recent times, the use of Artificial Intelligence (AI) has increased to address the governance challenges
facing cities. Due to its advanced capabilities, AI is expected to become a critical resource for local
governments in their pursuit of smart and sustainable development [
The application of chatbots in the public sector is not new [
GovAIEasy (Artificial Intelligence Makes It Easy) is an innovative AI-powered lexical simplification system employing the powerful GPT-4 human language generation model. Its main function is to convert complex texts into more accessible and easy-to-understand versions for users. This system is based on Lexical Simplification, with the aim of providing automatic summaries that improve users’ understanding of the text. Its approach aims to especially benefit people with low levels of literacy, cognitive problems, or disabilities that make it dificult to understand procedures in public ofices. By 1The GovAIEasy application is available at https://www.govaieasy.com/ transforming the complexity of language into a more understandable format, GovAIEasy contributes significantly to making state information more accessible and equitable for all. Figure 1 shows the graphic symbol that represents the GovAIEasy platform.
improve your understanding, the system automatically provides definitions, examples, and use cases for complex words, allowing us to illustrate the inner chain of thought (CoT) strategy followed by the model. This approach further facilitates the process of understanding the content for the user, providing them with the necessary tools to address specific dificulties that may arise when interacting with state documents, and providing a pedagogical way to improve reading abilities.
Our strategy focuses on applying zero-shot techniques to the GPT-4o model to generate accurate text sequences. Using the OpenAI API, the model is optimized to efectively meet the needs of GovAIEasy users, a solution that combines artificial intelligence to improve government documents and facilitate access to information in government. GovAIEasy ofers tools to analyze, clean and simplify texts, reducing complexity and improving comprehension (see Figure 2). It also uses GPT-4o for automated content analysis and summarization. Details of the features of the model are provided in Table 1. 3.1.1. Process Flow Description The stages considered in this study are fundamental to understanding the process of assessing lexical complexity in texts. Through a systematic approach, various phases are addressed that allow not only the identification of complex words but also their level of dificulty, generate accessible summaries, and store the results for later analysis. This comprehensive approach allows for a more precise assessment of the impact of complex words on text comprehension, which is essential for simplifying language in
3 presents the main stages involved in this process, described 1. Input data: The user provides a text. 2. Generate prompts: Two prompts are generated to send to the language model: • Prompt for complex words: The model identifies complex words and assigns a dificulty value.
• Prompt for summary: The model generates a concise summary of the text. 3. Processing by the model: The model processes both prompts and returns the results: identified complex words and summary.
a database and returned as a response to assess and study their lexical complexity, which will be used in future research. This storage enables not only to analyze the presence and distribution of dificult terms in the processed texts, but also to monitor the model’s performance in identifying and classifying these words.
Prompt design optimizes the way the model interacts with. The main advantage of this technique is that it does not require additional resources, as the performance of the existing model can be improved simply by adjusting the input presented to it [15]. Furthermore, it has been shown that proper prompt design in a general model, such as GPT-4, can outperform a model trained for a specific task [ 16] [17]. Since not all prompts generate the desired behavior in language models, it is essential to develop a prompting strategy that fits the task at hand. Although prompts often need customization for specific tasks, they can often be generalized within an appropriate prompt framework [ 18].
We define two main prompts that are sent to the language model to process the text: 1. Prompt for complex words. The first prompt aims to identify the most complex words in the text. The language model must extract a list of dificult words, assign them a dificulty value within the range of 0 to 1, and mark whether these words are acronyms. This process is carried out on the basis of instructions that establish diferent levels of dificulty (moderately dificult, dificult, and very dificult) according to the calculated complexity value.
the text in Spanish. This step allows obtaining a simplified version of the content that facilitates comprehension.
Our methodology generates a simplified version of the original text, transforming complex content into a more accessible version, as shown in Table 2. Our approach adapts to the most common vocabulary of users of ecuadorian public institutions, ensuring that the simplification preserves the naturalness of administrative language. From a linguistic perspective, the resulting text preserves the meaning of the original, ensuring the retention of key terms, such as proper nouns, within the legal framework. An example of this is the reference to “Impuesto a la renta” a specific tax with defined regulations, whose name is retained. Furthermore, the model optimizes the writing by avoiding redundant and irrelevant content, without requiring prior linguistic knowledge from the user. Table 3 presents a comparison between the original text and the simplified version, highlighting the preservation of key terms, linguistic clarity, and the elimination of irrelevant information, while Table 4 presents a comparison between the original text and the simplified version, highlighting structural modifications and readability improvements. Source Text ¿Quién debe pagar? La declaración de Impuesto a la Renta es obligatoria para todas las personas naturales, las sucesiones indivisas y las sociedades, nacionales o extranjeras, domiciliadas o no en el país, conforme los resultados de su actividad económica; aún cuando la totalidad de sus rentas estén constituidas por ingresos exentos, a excepción de: Los contribuyentes domiciliados en el exterior, que no tengan representante en el país y que exclusivamente tengan ingresos sujetos a retención en la fuente. Las personas naturales cuyos ingresos brutos durante el ejercicio fiscal no excedieren de la fracción básica no gravada. La normativa tributaria define a estos contribuyentes como “sujetos pasivos”.
Simplified Text La declaración de Impuesto a la Renta es obligatoria para personas naturales, sucesiones indivisas y sociedades, tanto nacionales como extranjeras, según su actividad económica, excepto para contribuyentes en el exterior sin representante y con ingresos sujetos a retención, y personas naturales con ingresos brutos que no superen la fracción básica no gravada.
Although CoT is considered the primary tool for reasoning in large-scale language models (LLMs), complex reasoning remains a major challenge for these models [19] [20]. Recent research has shown that explanations generated using CoT can be influenced by content biases, negatively afecting both their robustness and fidelity [ 21]. In Figure 4, our experiments demonstrated that the definitions, Original Sentence
Simplified Sentence
Lexical Change conforme los resultados de su actividad económica según económica su
actividad Los contribuyentes domiciliados en el exterior, que no tengan representante en el país y que exclusivamente tengan ingresos sujetos a retención en la fuente.
Las personas naturales cuyos ingresos brutos durante el ejercicio fiscal no excedieren de la fracción básica no gravada.
contribuyentes en el exterior sin representante y con ingresos sujetos a retención personas naturales con ingresos brutos que no superen la fracción básica no gravada.
“conforme los resultados de” → “según” (more concise) “domiciliados en el exterior” → “en el exterior”, “que no tengan representante en el país” → “sin representante”, “exclusivamente tengan ingresos sujetos a retención en la fuente” → “con ingresos sujetos a retención” “cuyos ingresos brutos durante el ejercicio fiscal no excedieren de” → “con ingresos brutos que no superen” Original Sentence ¿Quién debe pagar? (Interrogative sentence) La declaración de Impuesto a la Renta es obligatoria para todas las personas na-turales, las sucesiones indivisas y las sociedades, nacionales o extranjeras, domiciliadas o no en el país. aún cuando la totalidad de sus rentas estén constituidas por ingresos exentos, a excepción de: Los contribuyentes domiciliados en el exterior... + Las personas naturales cuyos ingresos brutos... (two separate sentences).
La normativa tributaria define a estos contribuyentes como “sujetos pasivos”
Simplified Sentence
(deleted) La declaración de Impuesto a la Renta es obligatoria para personas naturales, sucesiones indivisas y sociedades, tanto nacionales como extranjeras.
(deleted) excepto para contribuyentes en el exterior sin representante y con ingresos sujetos a retención, y personas naturales con ingresos brutos que no superen la fracción básica no gravada.
(deleted)
Structural Change The initial question was removed, as the simplification begins directly with the statement. “todas” and “domiciliadas o no en el país” were removed to make the sentence more concise.
This section was removed because it did not provide essential information for defining who is required to file a tax return.
Both exceptions were merged into a single sentence, making the structure more fluid.
The definition of “sujetos pasivos” was removed, as it is not essential to understanding the tax obligation. examples, use cases, and the complexity level of words identified as complex generated by GPT-4o through its CoT approach, provide accurate and consistent reasoning in the context of state texts. This CoT demonstrates the model’s relevance, as it processes and understands the input text through multiple levels of abstraction to produce a coherent and contextually relevant output. We highlight its efectiveness when evaluated on texts of varying complexity, achieving state-of-the-art performance and demonstrating consistent improvements in robustness.
The GovAIEasy application is designed to expand its application in the future to various types of text beyond public administration content. The system has a modular architecture that makes it easy to add or update features as needed. Periodic maintenance tasks are carried out on the collected data, as well as version updates to improve both the interface and the inclusion of new processes that enrich the analysis of the data in addition to performing regular error analysis to ensure performance and reliability precision of the automated system.
the improvement of public services specifically about attention and assistance to users. Governments can also use AI to improve communication with citizens, as well as to increase the eficiency and quality of public services [22]. This article relies on state-of-the-art systems, based on the premise that recent advances in artificial intelligence, particularly through the application of the Generative Pre-trained Transformer (GPT-4o) model, ofer a promising opportunity to transform government administration and enhance public services for the direct benefit of citizens.
It is important to highlight that storing complex words and their dificulty levels in a database not only allows for a detailed analysis of lexical complexity in texts, but also provides a key tool for evaluating model performance and improving its accuracy in future implementations. This information is essential for developing strategies that promote language accessibility and comprehension, especially in oficial and public interest documents. Thus, research in this field contributes to the creation of more efective strategies for text simplification, facilitating communication and ensuring that information is more understandable to a wider audience.
This work is funded by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project Desarrollo Modelos ALIA. This work has also been partially supported by Project CONSENSO (PID2021-122263OB-C21), Project MODERATES (TED2021-130145B-I00) and Project SocialTox (PDC2022-133146-C21) funded by MCIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR and Project ROMANET (CERV-2024-CHAR-LITI-101215052), funded by the European Union NextGenerationEU/PRTR.
During the preparation of this work, the authors used ChatGPT (OpenAI) to improve the writing in terms of clarity, coherence, and comprehensibility, and the free version of Grammarly—integrated with the online LaTeX editor Overleaf—for grammar and spelling correction. After using these tools, the authors manually reviewed and refined the content as needed and take full responsibility for the final version of the manuscript. between citizens and government through AI-guided chatbots, Government Information Quarterly 36 (2019) 358–367. URL: https://www.sciencedirect.com/science/article/pii/S0740624X17304008. doi:https://doi.org/10.1016/j.giq.2018.10.001. [14] M. d. P. Insapillo Fatama, Implementación de chatbot con inteligencia artificial para el mejoramiento del sistema helpdesk en el gobierno regional loreto, iquitos 2023 (2023). [15] J. Liu, D. Shen, Y. Zhang, B. Dolan, L. Carin, W. Chen, What Makes Good In-Context Examples for
GPT-3?, arXiv preprint arXiv:2101.06804 (2021). [16] H. Nori, Y. T. Lee, S. Zhang, D. Carignan, R. Edgar, N. Fusi, N. King, J. Larson, Y. Li, W. Liu, et al.,
arXiv preprint arXiv:2311.16452 (2023). [17] Y. Tao, O. Viberg, R. S. Baker, R. F. Kizilcec, Cultural bias and cultural alignment of large language models, PNAS nexus 3 (2024) pgae346. [18] E. Hedlin, L. Estling, J. Wong, C. Demmans Epp, O. Viberg, Got It! Prompting Readability Using
Machinery, New York, NY, USA, 2025, p. 115–125. URL: https://doi.org/10.1145/3706468.3706483. doi:10.1145/3706468.3706483. [19] J. Meadows, A. Freitas, Introduction to mathematical language processing: Informal proofs, word problems, and supporting tasks, Transactions of the Association for Computational Linguistics 11 (2023) 1162–1184. [20] C. He, R. Luo, Y. Bai, S. Hu, Z. L. Thai, J. Shen, J. Hu, X. Han, Y. Huang, Y. Zhang, et al., Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems, arXiv preprint arXiv:2402.14008 (2024). [21] M. Turpin, J. Michael, E. Perez, S. Bowman, Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting, Advances in Neural Information Processing Systems 36 (2023) 74952–74965. [22] J. Berryhill, K. K. Heang, R. Clogher, K. McBride, Hello, World: Artificial intelligence and its use in the public sector (2019).