We present Superlim, a multi-task NLP benchmark and analysis platform for evaluating Swedish language models, a counterpart to the English-language (Super)GLUE suite. We describe the dataset, the tasks, the leaderboard and report the baseline results yielded by a reference implementation. The tested models do not approach ceiling performance on any of the tasks, which suggests that Superlim is truly difficult, a desirable quality for a benchmark. We address methodological challenges, such as mitigating the Anglocentric bias when creating datasets for a less-resourced language; choosing the most appropriate measures; documenting the datasets and making the leaderboard convenient and transparent. We also highlight other potential usages of the dataset, such as, for instance, the evaluation of cross-lingual transfer learning.
The developments of deep learning in natural language processing (NLP) in recent years have resulted in an unprecedented amount of computational power and data required to train state-of-the-art NLP models. This makes lower-resource languages, such as Swedish, increasingly more reliant on language transfer effects from English since they do not have enough data to train separate monolingual models. In this study, we investigate whether there is any potential loss in English-Swedish language transfer by evaluating two types of language transfer on the GLUE/SweDiagnostics datasets and comparing between different linguistic phenomena. The results show that for an approach using machine translation for training there is no considerable loss in overall performance nor by any particular linguistic phenomena, while relying on pre-training of a multilingual model results in considerable loss in performance. This raises questions about the role of machine translation and the use of natural language inference (NLI) as well as parallel corpora for measuring English-Swedish language transfer.
Relative word importance is a key metric for natural language processing. In this work, we compare human and model relative word importance to investigate if pretrained neural language models focus on the same words as humans cross-lingually. We perform an extensive study using several importance metrics (gradient-based saliency and attention-based) in monolingual and multilingual models, including eye-tracking corpora from four languages (German, Dutch, English, and Russian). We find that gradient-based saliency, first-layer attention, and attention flow correlate strongly with human eye-tracking data across all four languages. We further analyze the role of word length and word frequency in determining relative importance and find that it strongly correlates with length and frequency, however, the mechanisms behind these non-linear relations remain elusive. We obtain a cross-lingual approximation of the similarity between human and computational language processing and insights into the usability of several importance metrics.