Teaching

Data privacy is a critical issue in NLP which limits data sharing and development of NLP tools to protect natural persons. While anonymisation, i.e., the removal of all Personally Identifiable Information (PII) from the data (“Anne” → “X”), or replacement with placeholders (“Anne” → “firstname”) are common and effective ways to ensure compliance with privacy laws, they render the source data illegible and thus unusable for tool development and downstream applications. MLMs present themselves as a particularly suitable, lightweight, and privacy-friendly tool for the task of predicting pseudonyms from natural language context (cloze-style). A thesis on this topic could work on, e.g.: (i) developing and improving evaluation metrics for natural-language pseudonyms, (ii) refining the prediction pipeline via, e.g., NER and coreference resolution, or (iii) exploring continued pre-training of MLMs using content-word focused training data to improve, e.g., POS accuracy. This thesis offers co-supervision with Maria Irena Szawerna, University of Gothenburg.

Grammatical gender, while challenging for language learners, is mastered easily by native speakers during language acquisition. Despite its widespread reputation of being arbitrary, lexical gender in German can be predicted fairly reliably via semantic, morphological, and phonological cues using boosting tree models, adding evidence to the regularity of gender assignment. A thesis in this area would explore whether LLMs also show signs of a generalised gender assignment system in German by testing their behaviour on nonce words.

The Last Translation Benchmark is an international initiative to collect translation examples that are currently unsolvable by LLMs, together with handwritten rules to evaluate on them automatically. A thesis could consist of contributing to this benchmark in a systematic way, for example by targeting a specific set of constructions, or a language that is so far underrepresented, or a semi-automatic approach to discovering potential new examples.

Although speech-text language models such as Qwen2-Audio and SeamlessM4T have shown remarkable abilities to process both speech and text across multiple languages, it remains unclear to what extent these models rely on morphosyntactic and prosodic cues when interpreting spoken language. Inspired by existing benchmarks for evaluating prosodic understanding, this thesis aims to develop a multilingual benchmark for systematically evaluating how speech language models process prosodic and morphosyntactic information across languages.

Recent work has discovered that in their lower layers, LLMs form an implicit vocabulary, merging or ‘detokenising’ subword tokens back into representations of complete words. While results suggest that LLMs do this not only for multi-token words, but also for multi-word expressions or more complex small phrases, the linguistic nature of these representations has not been studied. Extending previous methods to the study of the implicit vocabulary beyond words would form a contribution to the question of implicit learning of grammatical structures in LLMs.

Universal Dependencies (UD) is a framework for consistent annotation of grammar across human languages. The largest UD treebank of any language is UD-HDT, created via automatic conversion from the Hamburg Dependency Treebank and not actively maintained since 2017. A thesis could integrate additional pre-conversion data and fix known errors and inconsistencies. A M.Sc. thesis would additionally train UD parsers before and after the fixes and evaluate them on German treebanks to demonstrate the impact of the changes.

Universal Dependencies (UD) is a framework for consistent annotation of grammar across human languages. Most UD treebanks have not been actively maintained since creation, accumulating validation errors that limit their usefulness. If you read (natively or with reasonable comprehension) any of the languages in the UD validation report, a thesis could adopt that treebank, fix errors and warnings, and expand feature coverage or cross-treebank consistency. A M.Sc. thesis would validate improvements by training and comparing parsers before and after.

Within the broader category of intransitive verbs, unaccusative (e.g., You fall) and unergative (e.g., You resign) verbs have special syntactic behaviour relating to the semantic role of the subject, with is agentive for unaccusative verbs and patient-like for unergative verbs. These syntactic behaviours are not trivial to detect because in standard sentences, the distinction is only visible through the semantics of the subject. Previous work has made a first attempt at unsupervised discovery of both categories, and the thesis would attempt to approve on this with more modern methods, or possibly extend it to further languages where unaccusative and unergative verbs might have different properties.

Noun compounds provide an interesting test case for LLMs' understanding of complex noun semantics, as well as generalisation to novel compounds. Previous work has investigated this for compounds of two nouns. Compounds of more than two nouns are semantically far more complex and remain unexplored as a test case for LLMs. We might for example ask, is a child camel jockey slave a type of (child|camel|jockey|slave)? Data for this could e.g. come from the recently release Compound Branching Resource, or could be automatically collected from corpora and then hand-annotated. The thesis would lead to the creation of a (possibly multilingual) benchmark of LLMs' understanding of complex noun compounds and their internal structure.