Training Data Lab participates in the XII Latin American Congress of Political Science

From 17 to 20 July 2024, our research group, Training Data Lab, had the honour of participating in the XII Latin American Congress of Political Science, entitled “América Latina como actor en la configuración geopolítica global,” held in Lisbon, Portugal. Our research associates presented three papers reflecting our political science and methodology research.

Presentations

First, Carla Cisternas presented her paper ‘Expertise ad hoc: evidence from advisory commissions in Chile (1990-2022)’. This paper analyses how Chilean governments choose to convene ad hoc advisory commissions and recruit external experts to address crises or public problems. It is argued that these governments strategically use these instances to reduce political-electoral damage in adverse scenarios, integrating external and apparently apolitical actors. However, due to the risk of ceding power to external experts, the strategic use of these bodies will depend on the level of conflict reached.

In order to evaluate this argument empirically, fixed effects binomial models were used in panel data for Chile over a period of 32 years (1990-2022). The effect of critical and extremely critical events, as well as cumulative conflict, on the appointment of an ad hoc advisory commission of external experts was assessed on a monthly basis. Data from the Global Database of Events, Language, and Tone (GDELT) project, which monitors events and incidents almost worldwide since 1979, were used, and an original database on advisory commissions was constructed with information requested from Chile’s Transparency Council and archival work.

In second place, Bastián González-Bustamante presented two papers: “Peticiones de renuncia y dimisiones ministeriales en las democracias presidenciales latinoamericanas” and “Large Language Models (LLMs) para identificar toxicidad en la esfera digital durante eventos de protesta en América Latina”.

In the first paper, he analyses the effect of resignation calls and ministerial resignations on presidents’ protection policy and firing rule in 12 Latin American countries between 1976 and 2021. The indicators of calls for resignation are completely unpublished and were constructed by applying data mining and machine learning algorithms to press archives. Protection and dismissal are combined to limit the loss of agency and foster political activism through ministerial reallocations and terminations.

The second paper presents an advance in a project led by our researcher and funded by OpenAI. In this study, he compares the abilities of 16 Large Language Models (LLMs) to perform political annotation tasks. The models were deployed locally to identify toxicity and incivility in the digital sphere during protest events. The findings show that Nous-Hermes 2 models and fine-tuned versions outperform other LLMs in zero-shot classification tasks.

Conclusions

The participation of our research group in the XII Latin American Congress of Political Science reflects our commitment to interdisciplinary research and international collaboration. Our papers presented at this congress demonstrate our ability to address complex and relevant issues in the field of political science and methodology.

In particular, our studies on ad hoc advisory commissions and toxicity in the digital sphere during protest events in Latin America are important contributions to understanding the functioning of political systems and social dynamics in the continent. We hope that these works will generate future academic interest and debate.

* AI-generated text
Read more about how we generate our content