This article reports on an ongoing project aiming at automatization of pseudonymization of learner essays. The process includes three steps: identification of personal information in an unstructured text, labeling for a category, and pseudonymization. We experiment with rule-based methods for detection of 15 categories out of the suggested 19 (Megyesi et al., 2018) that we deem important and/or doable with automatic approaches. For the detection and labeling steps, we use resources covering personal names, geographic names, company and university names and others. For the pseudonymization step, we replace the item using another item of the same type from the above-mentioned resources. Evaluation of the detection and labeling steps are made on a set of manually anonymized essays. The results are promising and show that 89% of the personal information can be successfully identified in learner data, and annotated correctly with an inter-annotator agreement of 86% measured as Fleiss kappa and Krippendorff's alpha.

DaLAJ-GED is a dataset for linguistic acceptability judgments for Swedish, covering five head classes: lexical, morphological, syntactical, orthographical and punctuation. DaLAJ-GED is an extension of DaLAJ.v1 dataset (Volodina et al., 2021a,b). Both DaLAJ datasets are based on the SweLL-gold corpus (Volodina et al., 2019) and its correction annotation categories.DaLAJ-GED presented here contains 44,654 sentences, distributed (almost) equally between correct and incorrect ones and is primarily aimed at linguistic acceptability judgment task, but can also be used for other tasks related to grammatical error detection (GED) on a sentence level. DaLAJ-GED is included into the Swedish SuperLim 2.0 collection, 1 an extension of SuperLim (Adesam et al., 2020), a benchmark for Natural Language Understanding (NLU) tasks for Swedish. This paper gives a concise overview of the dataset and presents a few benchmark results for the task of linguistic acceptability, i.e. binary classification of sentences as either correct or incorrect.

We present our initial experiments on binary classification of sentences into linguistically correct versus incorrect ones in Swedish using the DaLAJ dataset (Volodina et al., 2021a). The nature of the task is bordering on linguistic acceptability judgments, on the one hand, and on grammatical error detection task, on the other. The experiments include models trained with different input features and on different variations of the training, validation, and test splits. We also analyze the results focusing on different error types and errors made on different proficiency levels. Apart from insights into which features and approaches work well for this task, we present first benchmark results on this dataset. The implementation is based on a bidirectional LSTM network and pretrained FastText embeddings, BERT embeddings, own word and character embeddings, as well as part-of-speech tags and dependency labels as input features. The best model used BERT embeddings and a training and validation set enriched with additional correct sentences. It reached an accuracy of 73% on one of three test sets used in the evaluation. These promising results illustrate that the data and format of DaLAJ make a valuable new resource for research in acceptability judgements in Swedish.

No abstract