Motivation:

LLMs alone are not reliable for critical decisions (e.g., due to the risk of “hallucinations” or incomplete information extraction). Ontologies and knowledge graphs can represent encyclopedic knowledge, but they often do not “communicate” with language models. It is therefore necessary to combine text (from which LLMs learn) with semantic structures (ontologies, knowledge graphs) to enable more equitable and evidence-based decision-making.

Methodology:

The ontology is grounded in two key theoretical frameworks. The first, Theory of Change (UN), provides a roadmap for understanding how an intervention can lead to concrete, measurable outcomes through evidence-based causal chains. The second, Nudging Theory (Thaler), emphasizes the use of subtle interventions that can guide people’s behaviors in predictable ways without imposing mandates or economic constraints.

• The ontology represents social and mobility data, including:
  • Geospatial, demographic, and mobility data: providing the spatial and population context needed for informed decision-making.
  • Competency-driven questions: for example, identifying which areas require improvements in public transport.
  • Integration of social perspectives and stakeholders: ensuring that multiple viewpoints are considered in the analysis.
  • Modeling of sensitive topics: such as caregiving, accessibility, and behavioral nudges.
  • Evidence from scientific literature: capturing research on citizens’ transport choices and relating it to demographic characteristics.

An example of the population in the knowledge ontology, extracted from scientific articles, is shown in the figure, illustrating the issue of spatial mismatch (Cui et al., 2022).

Here’s the different components of the Nudging Ontology:

Examples of represented information:

• Public transport policies
• Services and cost of living by area
• Indirect income estimates
• Data on age, gender, migration status, and household size
• Housing overcrowding and territorial inequalities

Below is an example of a query on the populated Knowledge Graph:

The Dataset:

• It contains 7,019 examples related to Turin, with data from 2012 to 2019.
• It covers 3,850 census areas, aggregated into 93 statistical zones and 9 districts.
• It has evolved from 16 to:
  • Demographics: population, gender, age, foreign residents, households.
  • Public transport: stops, lines, average distance.
  • Urban geography: areas, zones, districts.
  • Traffic and accidents: number, type, involvement of public transport.

Main Challenges:

• High complexity: many features need to be maintained without loss of information.
• High computational cost: 27 hours required to verbalize with the LLM.
• Need for more efficient verbalization schemes.

Verbalization Strategies:

1. Zero-shot: without guiding examples, the model generates freely (risk of errors).
2. Few-shot: with guiding examples (even just one), improving coherence and accuracy.
3. Structured (JSON): data representation without using the LLM, excellent for information retrieval.
4. Hybrid: combines the structured and few-shot approaches, with potential future integration of ontological rules.

In addition to the semantic integration of data, the project also explored new ways of representing and narrating information. In particular, heatmaps were created to describe aspects related to:

• The demographics of different areas of the city
• The distribution and accessibility of public transport
• Urban mobility patterns

These maps were not used solely as visual tools, but also as inputs for an image-to-text verbalization approach. The idea is that a language model can “read” the maps, describing their contents in textual form, and thereby generate coherent and interpretable narratives for policymakers or the general public.

The initial experiments have shown promising results: the generated descriptions not only reproduce the underlying data but also provide an accessible translation, making it easier to understand territorial disparities and urban dynamics. This paves the way for future developments in which text, structured data, and visual representations can be integrated into a single analysis and communication pipeline.