Migration is one of the main societal challenges faced by many countries in the European Union (EU). Thus, eficient and transparent processes to achieve the swift registration, health support, and social integration of Third Country Nationals (TCNs) is imperative. Although governments usually provide essential information through various oficial websites, these resources often lack multilingual accessibility. In some cases, critical content is available in the hosting country's native language. This language barrier can prevent migrants from accessing important information about public services, migration policies, and asylum procedures. This paper presents the development framework for the SALLY chatbot, designed to address some of these challenges by leveraging advancements in generative AI. The proposed solution integrates state-of-the-art technologies in Large Language Models, Dialogue Management, Knowledge Graphs, and Information Retrieval to facilitate personalized, multilingual interactions.
Third Country Nationals (TCNs) often face significant challenges integrating into host countries due to
diferences in education, culture, language, and legal status [
The SALLY project (Semantically Conscious Conversation-Based Chatbot Services for Migrants)
aims to develop an intelligent chatbot that assists migrants in Greece. Combining Large Language
Models [
Chatbot technology that supports individuals in diferent domains is a widely researched topic. For
example, CataractBot [
The author in [
In [
In another case, a chatbot [
Shifting to migrant support, the NADINE-bot [
In contrast to these works, the proposed chatbot focuses on the assistance of migrants in governmental related tasks and utilizes a plethora of state of the art technologies such as Large Language Models, Dialogue Management, Knowledge Graphs, and Information Retrieval.
The architecture of the SALLY chatbot has been designed to address the complex requirements of information delivery and interaction management for migrants in Greece. It combines multiple interconnected components that enable both natural language understanding and the delivery of precise and contextually appropriate information. Figure 1 provides an overview of the system’s architecture.
At the heart of the system is a Knowledge Base that organizes practical information related to migration procedures, services, contact points, facilities, and available benefits. To keep this data current, a dedicated web crawler regularly pulls updates from the oficial website of the Greek Ministry of Migration and Asylum. The collected content is grouped into five main hubs—procedures, benefits, services, facilities, and contacts—making it easier to retrieve the right information during conversations.
SALLY needs to be able to handle real-world conversations that may be informal, unclear, or multilingual. To make sense of user questions and generate helpful replies, SALLY uses Large Language Models (LLMs) such as Mistral, LLaMA 3, and GPT-4-Turbo. These models help interpret user input, generate relevant responses, and carry out tasks like detecting sentiment and identifying important entities.
To make sure that responses aren’t just fluent, but that they’re also accurate and up to date, the system utilizes the Retrieval-Augmented Generation (RAG) process. The RAG component is key to SALLY as it blends the broad language abilities of LLMs with factual information pulled directly from the knowledge base.
To keep conversations on track, a Dialogue Manager oversees the flow of interaction. It keeps track of context, user profiles, and past messages, so conversations stay coherent—even if they’re interrupted or continued after a break. It can also detect when a user seems confused or struggles with language, and adjust its responses to ofer clearer guidance.
SALLY also makes use of Semantic Conversational Spaces, which connect user data, dialogue history, and domain-specific knowledge through knowledge graphs. This setup helps the system understand relationships between diferent ideas and respond in a more informed, relevant way. External sources like DBpedia, ConceptNet, and BabelNet are used to enrich these spaces and support deeper understanding of the topics being discussed.
Finally, Moderation Filters are built into the system to help protect users from harmful or inappropriate content. These filters check for issues such as harassment, hate speech, violence, or privacy risks. SALLY combines advanced language tools, structured knowledge, and built-in safety measures to provide a helpful, respectful, and personalized experience for migrants in Greece seeking support.
The development of the SALLY chatbot highlights how combining knowledge-based systems with large language models can help meet information needs in complex and sensitive areas like migration. Thanks to LLMs, the system can understand natural, everyday language—even when it’s informal or influenced by diferent linguistic backgrounds, as is often the case with migrant users. At the same time, drawing on a well-structured knowledge base helps ensure that the information provided stays accurate and in line with oficial procedures. During development, several challenges came up. One of the biggest was keeping the generated responses consistent with verified information from the knowledge base. LLMs are good at sounding fluent, but sometimes they stray from the facts. Using a retrieval-augmented generation (RAG) setup helped reduce this issue by anchoring answers in real data, but more work is needed to fully control what the model produces.
Another challenge involved handling user profiles and keeping track of conversation history—especially when users return after long breaks or interact in unpredictable ways. The Dialogue Manager and Semantic Conversational Spaces helped keep things coherent across sessions, but there’s still room for improvement, particularly in personalizing the experience to each user. Privacy and safety were also top priorities, considering how vulnerable the target users may be. We used moderation iflters to prevent inappropriate or harmful content, but ongoing updates and monitoring are essential, especially as new issues or user behavior patterns show up. We also had to consider how SALLY would perform in real-world situations. Things like poor internet access, limited devices, or language-related obstacles can impact the user experience. While early testing has been positive under standard conditions, further trials with actual users will be needed in order to understand real-life constraints and make necessary adjustments.
The SALLY chatbot marks a meaningful step toward using generative AI to support migrant integration. By bringing together tools like Large Language Models, Dialogue Management, Knowledge Graphs, and Retrieval-Augmented Generation, it creates a conversational experience that feels personal, helpful, and tailored to the needs of migrants in Greece. Its support for both English and Greek allows for smooth communication, helping users find the information they need about services, procedures, and available support. At the same time, built-in safety measures promote respectful, constructive interactions—something especially important for vulnerable users.
Beyond its specific purpose, SALLY also shows how generative AI can help build more inclusive digital services. As more public and private organizations turn to AI-powered tools, systems like SALLY can serve as examples of how to design accessible, multilingual platforms that truly meet people where they are. By making it easier to access accurate information and get real-time help, SALLY not only assists individual users—it also supports better service delivery and stronger connections between migrants and their host communities.
This research project is carried out within the framework of H.F.R.I call “Basic research Financing (Horizontal support of all Sciences)” under the National Recovery and Resilience Plan “Greece 2.0” funded by the European Union - NextGenerationEU (H.F.R.I. Project Number: 15010).
During the preparation of this work, the authors used ChatGPT-4 for grammar and spelling checks. The authors have subsequently reviewed and edited the content and take full responsibility for the publication’s final version.