Automatic spell and grammar checking can be done using various system architectures, and large language models have recently been used to solve the task with promising results. Here we describe a new method of creating test data to measure the performance of spell and grammar checkers, including large language models. Three types of test data represent different approaches to evaluation, from basic error detection to error correction with natural language explanations of the corrections made and error severity scores, which is the main novelty of this approach. These additions are especially useful when evaluating large language models. We present a spell and grammar checking test set for Icelandic in which the described approach is applied. The data consists of whole texts instead of discrete sentences, which facilitates evaluating context awareness of models. The resulting test set can be used to compare different spell and grammar checkers and is published under permissive licenses.

Grammatical error correction (GEC) is the task of correcting typos, spelling, punctuation and grammatical issues in text. Approaching the problem as a sequence-to-sequence task, we compare the use of a common subword unit vocabulary and byte-level encoding. Initial synthetic training data is created using an error-generating pipeline, and used for finetuning two subword-level models and one byte-level model. Models are then finetuned further on hand-corrected error corpora, including texts written by children, university students, dyslexic and second-language writers, and evaluated over different error types and error origins. We show that a byte-level model enables higher correction quality than a subword approach, not only for simple spelling errors, but also for more complex semantic, stylistic and grammatical issues. In particular, initial training on synthetic corpora followed by finetuning on a relatively small parallel corpus of real-world errors helps the byte-level model correct a wide range of commonly occurring errors. Our experiments are run for the Icelandic language but should hold for other similar languages, and in particular to morphologically rich ones.

We train several language models for Icelandic, including IceBERT, that achieve state-of-the-art performance in a variety of downstream tasks, including part-of-speech tagging, named entity recognition, grammatical error detection and constituency parsing. To train the models we introduce a new corpus of Icelandic text, the Icelandic Common Crawl Corpus (IC3), a collection of high quality texts found online by targeting the Icelandic top-level-domain .is. Several other public data sources are also collected for a total of 16GB of Icelandic text. To enhance the evaluation of model performance and to raise the bar in baselines for Icelandic, we manually translate and adapt the WinoGrande commonsense reasoning dataset. Through these efforts we demonstrate that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus. We further show that initializing models using existing multilingual models can lead to state-of-the-art results for some downstream tasks.

Lemmatization, finding the basic morphological form of a word in a corpus, is an important step in many natural language processing tasks when working with morphologically rich languages. We describe and evaluate Nefnir, a new open source lemmatizer for Icelandic. Nefnir uses suffix substitution rules, derived from a large morphological database, to lemmatize tagged text. Evaluation shows that for correctly tagged text, Nefnir obtains an accuracy of 99.55%, and for text tagged with a PoS tagger, the accuracy obtained is 96.88%.

We report on work in progress which consists of annotating an Icelandic corpus for named entities (NEs) and using it for training a named entity recognizer based on a Bidirectional Long Short-Term Memory model. Currently, we have annotated 7,538 NEs appearing in the first 200,000 tokens of a 1 million token corpus, MIM-GOLD, originally developed for serving as a gold standard for part-of-speech tagging. Our best performing model, trained on this subset of MIM-GOLD, and enriched with external word embeddings, obtains an overall F1 score of 81.3% when categorizing NEs into the following four categories: persons, locations, organizations and miscellaneous. Our preliminary results are promising, especially given the fact that 80% of MIM-GOLD has not yet been used for training.