NLIP Seminar Series

NOTE THE UNUSUAL TIME FOR THIS SEMINAR Language models work well for a small number of languages. For the other languages, the best existing language model is likely multilingual, still with the vast majority of the training data coming from English and a few "priority" languages. We show that in many cases...

The alignment of large language models (LLMs) can often be brittle when faced with the complexities of real-world deployment. In this talk, I share our investigations on two scenarios where special care is required to ensure robust alignment. The first scenario is multi-objective alignment, where balancing competing...

The goal of Mechanistic Interpretability research is to explain how neural networks compute outputs in terms of their internal components. But how much progress has been made towards this goal? While a large amount of Mechanistic Interpretability research has been produced by academia, frontier AI companies such as Google...

Understanding and mitigating hallucinations in Large Language Models (LLMs) is crucial for ensuring reliable content generation. While previous research has primarily focused on “when” LLMs hallucinate, our work explains “why” and directly links model behaviour to the pre-training data that forms their prior knowledge...

Abstract: The interdisciplinary pilot project `Exploring novel figurative language to conceptualise Large Language Models’ is funded by Cambridge Language Sciences. This talk mainly concerns `slop', by which we mean text delivered to a reader which is of little or no value to them (or is even harmful) or is so verbose or...

Memorisation is a natural part of learning from real-world data: neural models pick up on atypical input-output combinations and store those training examples in their parameter space. That this happens is well-known, but which examples require memorisation and where in the millions (or billions) of parameters memorisation...

While model design gets much of the spotlight, subtle data choices, such as which documents are seen and how they’re represented, can profoundly shape the behaviour of language models. Nowadays, training data is the secret sauce behind a language model’s success, yet it remains relatively understudied. In this talk, I will...

To justify the generalisability of multilingual NLP research, multilingual language technology is frequently evaluated on ‘typologically diverse’ language selections. Yet, what this means often remains vague. In this talk, I first discuss what typological diversity means in NLP, and why it matters. Then, I introduce a...

Abstract: In this talk, I discuss benchmarking and interpreting large language models in the context of Swedish. I present a selection of work from my PhD thesis, which analyze LLMs Swedish-specific capabilities in different areas: English-Swedish language transfer, multi-task benchmarking on Swedish NLU and targeted...

Natural language processing used to rely on formal methods in its early days, and this included formal theories of syntax where sentence structure was of relevance. In the statistical era, the focus shifted to annotation schemes such as Penn Treebank and Universal Dependencies, which still rely on formal theory in their...