A Chatbot for Asylum-Seeking Migrants in Europe

Not surprisingly, migration is one of the top current concerns of the European Union too, and there’s every indication that the phenomenon is expanding. In 2023 there were almost one million asylum applications in the Member States of the European Union,⁴⁴4https://ec.europa.eu/eurostat/data/database a 21.3% increase compared to the previous year and the highest number since 2016. Because of recent conflicts, the number of irregular arrivals and asylum applications is estimated to grow further.⁵⁵5International Centre for Migration Policy Development (ICMPD), Migration Outlook 2024, https://www.icmpd.org/about-us/icmpd-migration-outlook. To handle the phenomenon more effectively and humanly [9], the EU Commission has set different reforms. Among them is the recent Pact on Migration and Asylum, i.e., a set of new rules managing migration and establishing a common asylum system at the EU level. Such a normative effort was made necessary by the complexity of current asylum rules. The application of such rules requires a non-trivial evaluation of legal and factual data concerning the applicant’s living conditions, both in the host country and in the country of origin, which makes it difficult for the territorial commissions, Courts, and even for migrants, to achieve a preliminary overview on their chances to obtain protection.

Taking all this into account, we believe it would be beneficial to have a tool for preliminary assessment and support, to orient the applicants and help them understand the possibilities and the requirements even before they begin the formal asylum request process. A tool that goes beyond the checklists used for handling well-defined, simple procedures since there is not only a problem of evaluating legal and factual data, but there is also an issue with understanding which procedures are relevant. Indeed, there is not only one type of protection but several ones. Importantly, since applicants may be political refugees and victims of abuse, discrimination, and persecution, the collection and processing of their personal data for immigration purposes must satisfy requirements of privacy; whereas the legal domain requires transparency and auditability concerning the application of the reference normative framework.

A possible approach to building such a tool could rely on computational argumentation. Recently there have been some proposals for argumentation-enabled chatbots supporting persuasive interaction [3] and privacy-preserving medical advice [6]. Argumentation is particularly suited to this sort of application due to its amenability to capture rich reasoning patterns and interconnections between alternative options and supporting elements, as well as dialogical interactions [14]. Computational argumentation frameworks are ideal candidates to support reasoning tools that must exhibit transparency and auditability requirement, because they are symbolic and thus interpretable [15]. The tool we propose is ACME: a chatbot for migrants in Europe. We envision ACME as a modular system, with several interfaces to fit different contexts. For example, for practical reasons, we believe a speech-to-text interface will be part of it.

Our work is partially based on our previous proposal of an information chatbot for COVID-19 vaccinations [6]. We adopted the same system architecture and preserved its properties, but we extended the chatbot to improve modularity and natural language interface, and to model a much more complex immigration case study than COVID-19 vaccines. Some important benefits of the ACME chatbot are its properties towards satisfying the domain requirements. Data governance and privacy are preserved through the modular architecture; transparency and explainability are guaranteed by the use of argumentative reasoning; the explicit, formally-defined knowledge base and reasoning methods are expert-made and fully auditable.

Finally, as a disclaimer, we shall stress that the purpose of our tool is to support, inform, and guide asylum applicants, not to assist or replace a judiciary expert in a “predictive justice” fashion, a topic that raises several ethical concerns [12] and is outside of our scope.

Conversational agents can benefit from computational argumentation in terms of explainability and auditability, improving user trust, and exploiting a reasoning engine to guide the conversation. The application of argumentative chatbots spans several topics such as healthcare [1, 17] and COVID-19 [3, 6], ethics [10], university fees [2, 8] and education [7]. Chalaguine and Hunter utilize a crowd-sourced argument graph as a knowledge base to support a chatbot persuading the user, considering the UK’s university fees [2] and COVID-19 vaccines [3] as case studies. The conversational agents described in these studies address the user’s concerns by associating them with an argument in the graph and replying with its counterargument. Aiming at broadening the user’s viewpoint, Hauptmann et al. [10] have developed a German argumentative chatbot to converse about AI-related ethically sensitive topics. The system determines the user’s stance on a certain topic and challenges it. In their study to implement a healthcare chatbot, Castagna et al. [1] employ an Explanation-Question-Reply (ERQ) argument scheme modelling the agents’ interactions and clinically specialized argument schemes (ASPT). The chatbot aims to deliver patients’ treatment plans and answer the user’s questions. As far as conversational agents in the immigration domain are concerned, Kotiyal et al. [11] leverage a knowledge graph containing information about the UK’s immigration laws to aid the user in preparing a visa request using question templates as a language interface, but does not use argumentation in any way.

The ACME architecture is inspired to our previous work [6, 5, 4]. We adopt a “neuro-symbolic” architecture, where a neural module dedicated to understanding the user is combined with a symbolic module dedicated to reasoning. ACME’s core components are:

• •

  a knowledge base (KB) consisting of a text file representing an argumentation graph and a JSON file that associates a set of natural language sentences with each node;

• •

  a neural natural language module whose task is to interact with the user, understanding what they write;

• •

  a symbolic argumentation module to reason from all the available data, compute answers and provide explanations.

The user interacts with the natural language module providing relevant information and answering questions. The information thus elaborated is mapped onto concepts of the KB. The argumentation module performs reasoning on the KB and finds the appropriate answer. In case a definite answer is not possible, the system identifies a node that would help determine an answer and asks the user to provide such information.

The architecture design is focused on decoupling between the natural language module and the argumentation module. The latter has access only to the information mapped into the concepts of the KB, which by definition are relevant for inferring an answer. Any other irrelevant information provided by the user is discarded, guaranteeing their privacy.

The Refugee Status is the strongest level of protection that can be granted. It is given to anyone who fears persecution for reasons of race, religion, nationality, membership of a particular social group or political opinion and is unable or unwilling to avail himself of the protection of his country. Subsidiary Protection Status is a second-level strong guarantee which takes place mostly for the same reasons as the refugee status but only if the latter cannot be recognized. Given the similarity of the requirements, for the purpose of this case study, we consider them to be the same. Finally, Special (or Humanitarian) Protection is a National type of protection; this is a weaker safeguard that is guaranteed to applicants who are not eligible as refugees or beneficiaries of subsidiary protection but worthy of protection due to special humanitarian reasons. Our case study investigates the possibility of applying for international protection in Italy by people whose country of origin is Nigeria. Our tool considers the conditions required by the two protection classes: the Refugee/Subsidiary Protection Status, and the Special Protection.

In most cases, migrants do not know the regulatory framework of international protection. Therefore, they are often unable to identify which kind of protection they can apply for. Most of the time, they apply for all types although they can only get one. ACME is meant to help the applicant understand which is the highest level of protection they can apply for.

It has often been observed that legal knowledge is difficult, if not impossible, to access for the individuals who should benefit from it [13]. The point has been made in particular for consumer law, but it also holds for the regulatory framework of international protection: a system whose beneficiaries, the migrants, are usually unable to navigate alone. In this work we described the demonstration of a system, ACME, intended to facilitate a sustainable approach to migration by guiding and supporting a migrant or someone acting on their behalf towards understanding which is the highest level of protection they can apply for. In its present realization, ACME is only a proof-of-concept prototype. However, such a prototype is needed in order to investigate the suitability and potential of the interplay between computational argumentation and natural language processing in such a complex domain. The overarching objective of this effort is to achieve the vision of a Human-Centered AI applied to a crucial global challenge. We hope that our work can inspire and motivate further research on this domain.

The main challenges we found so far in our endeavor had to do with addressing the complexities of a regulatory framework together with those of natural language understanding and the field requirements imposed by the target domain. We believe that the only way to success is through a collaboration of teams composed of legal, AI and domain experts from relevant organizations. In the future, we plan to extend our proof-of-concept to cover a larger variety of situations and further modules, such as a speech-to-text interface and automatic translation. We also aim to validate ACME with the legal experts involved in the asylum requests, such as lawyers and judges.

https://youtu.be/P8iW7FOZTYM

https://github.com/lt-nlp-lab-unibo/ACME-A-Chatbot-for-Migrants-in-Europe

We thank Avv. Mariapia Frisina for the valuable insights and discussions. This work was partially supported by the following projects: European Commission’s NextGeneration EU programme, PNRR – M4C2 – Inv. 1.3, PE00000013 - “FAIR - Future Artificial Intelligence Research” – Spoke 8 “Pervasive AI”; EU H2020’s ICT48 project “Humane AI Net" Grant #952026; MUR project PRIN 2022 EPICA "Enhancing Public Interest Communication with Argumentation" (CUP H53D23003660006) funded by Next Generation EU.