Projects (PI and co-PI)
Summary: The European Union (EU) faces pressing demands to act in major policy areas amid public contestation of supranational governance. Our interdisciplinary project seeks to explain and facilitate responsive and effective policy reforms by increasing knowledge about the willingness and capacity for EU integration in specific policy areas. We study the conditions under which the EU institutions seek to increase or decrease EU policy competences, when their positions respond to public demands across and within member states, and under what conditions each institution manages to assert its position in the policy-making processes. We further investigate how the institutions’ positions and capacity to steer the course of European integration across policy areas are reshaped by increased EU politicization and associated shifts in institutional identities, internal disunity and switch from formal political to informal technocratic procedures of policy-making. Public opinion surveys, party manifestos and speech data from European and national parliaments will serve to capture citizen, party and government preferences over the transfer of competences from the national to the EU level across policy areas. We will then examine under what conditions and to what extent these preferences determine the positions of EU institutions, policy proposals and adopted legislation with respect to the level of competence transfer to the EU, using cutting-edge methods for computational text analysis. These findings will serve to develop recommendations about innovation in policy and institutional design that can address pressing challenges and enjoy public acceptance in member states and among their citizens.
Goal: Bring state-of-the-art methods for multilingual language understanding to challenging problems in quantitative political science
Funding Body/Scheme: NORFACE Governance
Duration: Oct 2020-Sep 2023
Budget: 1.2 million EUR (ca. 350 thousand EUR for Uni Mannheim)
Multi2ConvAI: Multilingual and Multi-Domain Conversational AI
Summary: Conversational AI systems suffer from limited portability across domains and languages. across domains and languages. In this project, we aim to bring cutting-edge developments from the field of natural language processing language processing, namely multilingual representation learning and transfer learning, to enable faster and more faster and more cost-effective bootstrapping of conversational assistants for new domains and languages. languages. As a proof-of-concept, we will develop conversational models for several complex domains (e.g., supporting mechanics during vehicle inspections) and languages that are not supported in most commercial systems (e.g., Turkish, Hindi, Romanian, Tamil), relying on state-of-the-art domain state-of-the-art methods for domain transfer (e.g., adapter-based finetuning) and language transfer (e.g., multilingual (e.g., multilingual transformer models). The project will provide a platform for the creation and evalution of dialog systems and a repository of reusable, pre-trained dialog models.
Goal: modular approach to task-oriented dialog systems, suitable for real-world use cases of our project partners (Neohelden and inovex). Independent and recombinable dialog modules for different tasks, languages, and domains.
Funding Body/Scheme: Baden-Württemberg Wirtschaftsministerium, Program "AI made in Baden-Württemberg"
Duration: Mar 2018-Sep 2021
Budget: 450 thousand EUR (150 thousand for Uni Mannheim)
AGREE: Algebraic Reasoning over Events from Text and External Knowledge
Summary: In many domains of human activity (e.g., business or medicine) very valuable new knowledge can be inferred by combining novel events and developments (e.g., published on the web or in research articles) with the large body of existing (structured and unstructured) domain knowledge. This research project will investigate a framework for automatic induction of new knowledge upon discovery of novel domain information. We will extend our prior work on event extraction from text and combine it with the most recent developments in representation learning and natural language inference in order to support complex reasoning -- leading to induction of novel, non-trivial domain knowledge and prediction of future domain events.
Goal: Combine knowledge from text and external knowledge bases and ontologies for better automatic in-domain reasoning
Funding Body/Scheme: Baden-Württemberg Stiftung, Eliteprogramm
Duration: Mar 2018-Sep 2021
Budget: 142 thousand EUR