MULTILINGUALITY AND LANGUAGE TECHNOLOGY
Your Company Address
Deutsches Forschungszentrum für Künstliche Intelligenz GmbH (DFKI) Campus D3 2 Stuhlsatzenhausweg 3 66123 Saarbrücken Germany
The goal of this project is to generate a concise description of a rescue operation/disaster response mission as it is usually done in the mission reports but using only the mission dialogues (i.e., transcriptions) and some report templates as input.
Related projects: ADRZ, Cora4NLP
Potential supervisor: Tatiana Anikina
(Email if interested)
Prerequisites: good knowledge of German, experience with ML and neural networks; previous experience with NLG is an advantage.
BSc / MSc: can be either Bachelor or Master thesis topic
In this project, we will investigate different methods to transfer knowledge for Slot Filling (or NER as a closely related task) between languages, using English as the source language and German as the target one. We will utilize multiple open-source datasets for Slot Filling and NER available in English to facilitate training of neural models which will be applied to a low-resource domain in German: emergency call dialogues. Possible methods to investigate are: a) translation of the English data into German for model training; b) utilizing adapters (see Pfeiffer et al., 2020): task-specific adapters to learn the task on the training data in one language, language-specific adapters to switch between languages. You are invited to suggest and test any further approaches for the cross-lingual knowledge transfer. The results will be compared with each other to identify the most advantageous strategy. Investigating effect of different strategies for diverse model architectures is another possible research direction (we are currently working with Transformer-based architectures as well as RNN-based encoder-decoder models). You will work with simulated emergency call dialogues (112), identifying information crucial for the disposition of the forces. You will need to deal with constraints of a real-world application: the final system has to be time and memory efficient, it has to work in real time.
Related project: NotAs.
Potential supervisor: Anastasiia Kysliak
(Email if interested)
Prerequisites: good knowledge of German and English, good programming skills, familiarity with neural networks and PyTorch. Familiarity with Transformers library from HuggingFace is a plus.
BSc / MSc: can be a Master thesis topic.
Other topics are possible upon mutual agreement with the supervisor(s).
If you are interested in pursuing any of the topics, please send an email to the respective potential supervisor(s) listed for the respective topic, including a statement of interest, a short CV and a current transcript of records (as one PDF). As the number of places is limited, please apply by the deadline of March 4, 2022 for starting a thesis project during the Summer Semester 2022. We are looking forward to hearing from you!