Zero-shot Named Entity Recognition (NER) has gained prominence for information extraction across diverse domains without being limited to a single, fixed tag set. However, existing NER resources vary widely in data format, licensing terms, annotation schemes, and availability, making it difficult to systematically evaluate the generalization capabilities of zero-shot NER models. Prior attempts to aggregate datasets with broad coverage across domains have largely focused on a small subset of languages, and it is often not transparent how datasets were processed from their sources. This paper introduces MELD, a comprehensive multilingual and multi-domain data collection designed to address these gaps. MELD integrates 60 NER datasets spanning 194 languages, 14 domains, and 601 normalized entity types. While previously introduced multilingual NER datasets are mainly silver-standard, MELD contains gold-standard annotations for 60 languages. All data processing steps are fully open-source and reproducible, facilitating future extensions and ensuring long-term accessibility. While MELD is primarily designed for zero-shot evaluation, it also provides training and development splits in a single, consistent format to support future research in few-shot and supervised NER settings.

Training high-performing language models which include medium- and lower-resource languages remains a challenge. Massively multilingual models still underperform language-specific adaptations, especially at smaller scales, due to the curse of multilinguality. In this work, we investigate scaling as an efficient strategy for adapting pretrained models to new target languages. Through comprehensive ablations matched for data and training FLOPs, we test whether upscaling from English base models enables more effective and resource-efficient adaptation than standard continued pretraining. We find that upscaling is an efficient strategy for improving target-language performance by relieving capacity constraints, especially for higher-resourced languages. An initial upscaling step in English is beneficial for low-resource languages. However, in higher-resource scenarios it is more compute-efficient to directly scale up on the target language, albeit slightly compromising data efficiency for most languages. For multilingual continued pre-training, we identify directly scaling up using the target data as the optimal setup for both data efficiency and reduced English forgetting.

In visual communication, the ability of a short piece of text to catch someone’s eye in a single glance or from a distance is of paramount importance. In our approach to the SemEval-2020 task “Emphasis Selection For Written Text in Visual Media”, we use contextualized word representations from a pretrained model of the state-of-the-art BERT architecture together with a stacked bidirectional GRU network to predict token-level emphasis probabilities. For tackling low inter-annotator agreement in the dataset, we attempt to model multiple annotators jointly by introducing initialization with agreement dependent noise to a crowd layer architecture. We found our approach to both perform substantially better than initialization with identities for this purpose and to outperform a baseline trained with token level majority voting. Our submission system reaches substantially higher Match m on the development set than the task baseline (0.779), but only slightly outperforms the test set baseline (0.754) using a three model ensemble.

Our submission for Task 1 ‘Cross-lingual Semantic Parsing with UCCA’ at SemEval-2018 is a feed-forward neural network that builds upon an existing state-of-the-art transition-based directed acyclic graph parser. We replace most of its features by deep contextualized word embeddings and introduce an approximation to represent non-terminal nodes in the graph as an aggregation of their terminal children. We further demonstrate how augmenting data using the baseline systems provides a consistent advantage in all open submission tracks. We submitted results to all open tracks (English, in- and out-of-domain, German in-domain and French in-domain, low-resource). Our system achieves competitive performance in all settings besides the French, where we did not augment the data. Post-evaluation experiments showed that data augmentation is especially crucial in this setting.