ECPR

Install the app

Install this application on your home screen for quick and easy access when you’re on the go.

Just tap Share then “Add to Home Screen”

Analysing Arabic Interviews and Documents with Large Language Models

Methods
Qualitative
Quantitative
M2

Thursday 12:00 - 13:30 GMT (18/02/2027)

Abstract

Speakers: Steffi Dornschneider-Elkink, University College Dublin Johan A. Dornschneider-Elkink, University College Dublin Abess Benissmail, University of Galway Julius Strunk, Scuola Normale Superiore and the Sant’Anna School of Advanced Studies Chair: Irene Weipert-Fenner, Peace Research Institute Frankfurt In the seminar, two presentations demonstrate and discuss how large language models (LLMs) can support the analysis of Arabic-language material, and where they need careful guidance. While building on different empirical cases (details below), both presentations will discuss the advantages, challenges and pitfalls of using LLMs for interview data and large text corpus respectively. Abstracts of presentations: 1. Arabic interviews and LLMs: The first presentation by Steffi and Johan Dornschneider-Elkink and Abess Benissmail deals with interviews, but coding them by hand is slow, subjective and hard to replicate. Building on a study of female members of Hezbollah and drawing on 59 interviews conducted in and around Hezbollah in Dahiyeh and South Lebanon in 2022–23, and on eight books used to educate members of Hezbollah’s women’s branch, the presenters first explain why standard computational tools struggle with Arabic. In a second step, they present a pipeline built around an open-weight LLM that runs locally, so sensitive interview data never leaves the researchers' machines. The pipeline extracts causal links, codes them by category, life domain and stance, reconciles labels across texts, and produces a cognitive map for each interviewee. 2. Arabic documents and LLMs: Julius Strunk, in his work, examines how far parliamentary performance matters for elite retention in authoritarian states. It draws on a unique dataset on all parliamentarians in Jordan since the country’s independence in 1947 as well as Jordanian parliamentary minutes. In his presentation, he elaborates on the use of LLMs to manage this large amount of textual data. He focuses on data management – structuring minutes, quality control and mitigation, extracting relevant information – and analysis – political scaling. The first part mainly relies on Claude Code (Opus 5.5) and is code-based, whereas the second part mainly relies on the use of localised smaller LLM models – like Qwen 3.5.