The EPICA (Empowering Public Interest Communication with Argumentation) project aims to improve Public Interest Communication (PIC) through the use of advanced Computational Argumentation (CA) techniques. This paper provides a summary of the technical results of the project and discusses future research direction. In an era where digital communication rapidly shapes public opinion and social discourse, it is of paramount importance to ensure the efectiveness of Public Interest Communication (PIC) [ 1], which is used by institutions and organizations to address significant societal challenges and, in particular, to promote population health and well-being. The EPICA project (Empowering Public Interest Communication with Argumentation) started from the observation that the development of formal models and the use of information technology to support PIC activities are quite limited. It aims to integrate Computational Argumentation (CA) [2, 3] techniques into PIC strategies with the goal of providing innovative insights and tools to refine institutional communication practices. This paper provides a synthetic report of the activities carried out in the project up to the current date. In particular, Section 2 describes the main case study selected for the project, while Sections 3 to 6 summarize the result of the main Work Packages (WPs) of the project and Section 7 concludes.
As a use case, the EPICA project focused on public interest campaigns aimed at increasing fruit and
vegetable consumption, following guidelines largely established by the World Health Organization [
Targeted questionnaires have uncovered the presence of strong, yet largely unspoken,
counterarguments that influence behaviour without necessarily afecting individuals’ beliefs [
This third component reflects non-attitudinal barriers that a well-crafted message can address. In the context of greener diets, such barriers include the cost of fruits and vegetables, taste preferences, and the efort required to cook vegetables. These factors create incoherence: people may believe in the benefits of fruits and vegetables but still fail to consume the recommended amount.
Because these counterarguments rarely emerge in public discourse, where the benefits of fruit and vegetable consumption are typically taken for granted, we argue that they should be explicitly identified and addressed during the planning phase of any public interest communication campaign on the topic.
The first work package of the project has been devoted to draw a comprehensive conceptual analysis
to identify elements of institutional PIC activities that can benefit from advanced CA techniques and
to define requirements for the formalization of advanced argumentation models supporting PIC. The
analysis has considered a sample of 5 initiatives devoted to promoting greener diet, which have been
selected to ensure a variety in both the status of the promoting institutions, ranging from a national health
authority to not-for-profit organizations and professional dietitians working for private companies,
and in the forms of communication adopted, including institutional websites, a publicly available book,
and Top 10 lists of reasons to eat more fruits and vegetables. The contents of the initiatives have been
analyzed and compared with reference to the following aspects: the values (e.g., preservation of health,
environment protection, sensory pleasure, etc.) promoted by the arguments in the campaigns; the
argumentation schemes [
The second work package of the project has built on the analysis carried out in WP1. The requirements have been prioritised taking into account both their importance and their feasibility within the time span of the project. Accordingly, the activities have been focused on two main aspects: • first, the single audience scenario has been addressed, and a new concept of “robustness” has been defined. Specifically, the robustness of arguments and sets of arguments is considered from the point of view of the resistance of the accepted and of the extension status of arguments and sets of arguments, respectively, to changes in the audience preferences. • then, the scenario of diverse audiences has been considered. The focus in this context has been on developing a mathematical model to quantify and analyse how diferent inferential and epistemic standards, as well as the influence of non-inferential components like promoted values, afect the assessment of public interest communication among varied target audiences.
As to the first scenario, a new paradigm for reasoning on the robustness of the status of (sets of) arguments in (single-audience) value-based argumentation frameworks has been introduced. In particular, we investigated the decision problems underlying the computation of the degree of robustness, and thoroughly investigated their computational complexity. The novelty of the contribution lies in the fact that robustness has never been studied in the literature of abstract argumentation with regard to audience profiling, but only, indirectly, within the framework of enforcement, which however did not consider some features which are crucial in our context. In particular, it does not take into account values and preferences and is focused on turning non-acceptance into acceptance, while the opposite problem of turning acceptance into non-acceptance is also crucial for robustness.
As to the second scenario, a vector-based extension of value-based argumentation has been proposed,
introducing the concept of value vectors to represent a multi-dimensional spectrum of values influencing
audience perception and response. A notion of impact measure has then been introduced, which aims to
represent how combinations of weighted values associated with an argument are more or less efective
with respect to the attitudes of diferent audiences, also represented as vectors of weights. On this
basis, alternative notions of convincing arguments have been examined, referring either to the value
of the impact measure per se or to how it afects the acceptability of arguments, evaluated using an
argumentation semantics in the context of an argumentation framework [
The activities carried out in WP2 are described in the relevant deliverable [
We created a specialized corpus focused on campaigns promoting greener diets, with particular attention to increasing fruit and vegetable consumption. The corpus contains materials in both English and Italian. English-language campaigns were sourced from countries such as the United States, United Kingdom, Ireland, Canada, Australia, and New Zealand. Italian-language campaigns were exclusively drawn from initiatives targeting audiences in Italy. Data collection took place over a two-month period, from October to December 2024.
The dataset was built using a two-step methodology. First, we performed a manual Google search using queries like “fruit & vegetables campaign + country,” reviewing results from the first 15 pages. This was supplemented by ChatGPT searches, using prompts such as “Provide campaigns on fruit and vegetable consumption in + country” to identify additional content.
To ensure relevance and consistency, we applied strict inclusion criteria: just campaigns explicitly aimed at promoting fruit and vegetable consumption were selected. Broader public health campaigns (e.g., those focused on reducing salt intake or encouraging physical activity) were excluded. The final corpus includes 45 campaigns in total, 41 in English and 4 in Italian.
For each campaign with a dedicated website, all textual content was automatically extracted and subsequently processed through a pipeline designed to identify the minimal components of the argumentation.
We propose a pipeline for automatically inducing argument maps from text. The pipeline reframes traditional argument mining by focusing on textual entailment between extracted claims, rather than relying solely on sequential identification–classification–relation labeling steps. This methodological choice was driven by the fact that, at present, there is no argument mining pipeline capable of identifying the basic argumentative units in a text, classifying them by type, and determining the relations that exist between them. The pipeline is composed of four steps: • Claim Extraction: sentences are segmented and processed by a pre-trained model to extract elementary claims; • Co-reference Resolution: extracted claims are disambiguated to resolve pronoun references and contextual ambiguity; • Textual Entailment Analysis: all sentence pairs are evaluated using an LLM trained on Natural
Language Inference (NLI) tasks to detect entailment, contradiction, or neutrality; • Graph Construction: sentences and their entailment relations form a directed graph. The root of the largest connected component is treated as the major claim. Nodes connected to it are interpreted as supporting claims or premises, based on their structural positions.
This approach, through processing of PIC campaigns content, aims at producing graph-based summaries of argumentative content that work even for partially argumentative texts and can be increased with arguments from other texts.
Building on the foundational work carried out in WP2 and WP3, we focused on implementing the proposed reasoning algorithms and developing validation tools to present the approach to a selected group of users. In particular, starting from the work in [11] (the second scenario described in Section 4), we provided implementations for reasoning in the context of the vector extension to value-based argumentation. The tool we have developed1 is designed to support the research and analysis in the context of public campaigns. It enables the measurement of the efectiveness of public interest communication by analysing arguments with respect to multiple values and audiences. Arguments are associated with value vectors that represent their afinity to various factors influencing audience perception and response. By examining these value dimensions, the tool provides insight into the impact of communication strategies on diferent audience groups.
Interaction with the tool takes place through a web application and is structured into three steps: campaign definition , visualisation, and analysis. The campaign definition step allows users to model a campaign of public interest. The user can upload or paste a JSON file describing the entire campaign or, alternatively, use a dedicated form to enter each individual component. These components include positive arguments, arguments, attacks, audiences, and values associated with both arguments and audiences.
The visualisation step allows users to explore key aspects of the campaign. First, the tool shows the impact measure of each argument for diferent audiences. It then presents the computed defeat relation (↠ ) among arguments for all audiences, highlighting which attacks loose efectiveness with respect to the initial framework. Finally, the grounded semantics is computed and displayed for the framework ⟨, ↠ ⟩. The resulting set of arguments represents those considered convincing within the context of the campaign. The web page also includes an interactive graph that represents the campaign and allows modification by adding or removing arguments and attacks.
The last step is the analysis one, which implements two goal functions to identify the most efective positive arguments in the campaign for convincing the audience. The first objective, “Overall Efectiveness”, seeks the argument that maximises the weighted sum of its impact across all audiences. The second objective, “Convinced People”, identifies the argument that maximises the weighted number of audiences it successfully convinces. The scores of each argument with respect to the two goals are displayed in a table.
The EPICA project is at an advanced stage of development. It first developed a broad conceptual analysis and defined an extensive set of requirements concerning the use of computational argumentation techniques to support Public Interest Communication, with a special focus on the selected case study namely the promotion of a greener diet. Based on a selected set of the identified requirements, formal models have been developed considering issues like robustness and efectiveness of communication campaigns, with reference to scenarios involving both single and multiple audiences. To support the analysis and the validation of the proposed models in the context of the selected case study, a corpus focused on campaigns promoting greener diets has been developed. It results from an activity of data collection concerning campaigns carried out in a variety of countries worldwide, followed by an automated extraction phase, based on a suitable adapted mining pipeline. Finally, a prototypical tool implementing the proposed reasoning algorithms and supporting the selection of the most suitable arguments for a given campaign has been developed. The tool is available online and allows user to introduce and modify interactively the design of their communication initiatives. Among the future work directions, we mention further developments and refinements of the formal models, their validation using the developed dataset, the experimentation of the implemented tool with communication experts, and the dissemination of project results through events directed to relevant communities of stakeholders.
This work has been supported by MUR project PRIN 2022 EPICA (Master CUP D53D23008860006, unit CUPs J53D23007220006, H53D23003660006, B53D23013230006), funded by the European Union - Next Generation EU, mission 4, component 2, investment 1.1.
[11] P. Baroni, G. Fellin, M. Giacomin, C. Proietti, A vector-based extension of value-based argumentation for public interest communication, in: C. Proietti, C. Taticchi (Eds.), Proc, of the 8th Workshop on Advances in Argumentation in Artificial Intelligence, volume 3871 of CEUR Workshop Proceedings, CEUR-WS.org, 2024.