Web-crawled corpora are essential resources for linguistic and NLP research, offering far more data than is available from curated corpora. However, they often contain a great deal of low-quality texts which can complicate research and degrade the quality of pre-trained language models. Therefore, they are typically filtered, e.g. by applying rules or classifiers. In this paper, we compare the effectiveness of various text filtering classifiers and measure their impact on language model performance for three medium-resource languages. We present TQ-IS, an Icelandic text quality dataset consisting of 2,000 web-crawled documents, in which spans of low-quality text have been manually identified and labeled. We then evaluate a perplexity-based classifier, a supervised classifier trained on TQ-IS, and a self-supervised classifier trained to discern between documents from curated and web-crawled corpora on Icelandic, Estonian and Basque. We find that these classifiers obtain F1 scores of 94.48%, 99.01% and 93.40%, respectively, when evaluated on the TQ-IS dataset. Furthermore, our results show that while adding filtered web-crawled text to a pre-training corpus can improve downstream performance for pre-trained language models, any improvement is likely to remain modest unless the web-crawled corpus is significantly larger in size.

We experiment with sentiment classification models for Icelandic that leverage machine-translated data for training. Since no large sentiment dataset exists for Icelandic, we translate 50,000 English IMDb reviews, classified either as positive or negative, into Icelandic using two services: Google Translate and GreynirTranslate. After machine translation, we assess whether the sentiment of the source language text is retained in the target language. Moreover, we evaluate the accuracy of the sentiment classifiers on non-translated Icelandic text.The performance of three types of baseline classifiers is compared, i.e., Support Vector Machines, Logistic Regression and Naive Bayes, when trained on translated data generated by either translation service. Furthermore, we fine-tune and evaluate three pre-trained transformer-based models, RoBERTa, IceBERT and ELECTRA, on both the original English texts and the translated texts. Our results indicate that the transformer models perform better than the baseline classifiers on all datasets. Moreover, our evaluation shows that the transformer models trained on data translated from English reviews can be used to effectively classify sentiment on non-translated Icelandic movie reviews.

Web-crawled corpora offer an abundant source of training data for language models. However, they are generally noisy and are typically filtered using heuristic rules or classifiers. These methods require careful tuning or labeling by fluent speakers. In this paper, we assess the effectiveness of commonly applied rules on TQ-IS, a manually labeled text quality dataset for Icelandic. Additionally, we advocate for the utilization of unsupervised clustering and outlier detection algorithms for filtering. These algorithms are language-independent, computationally efficient and do not require language expertise. Using grid search, we find the optimal configuration for every combination of rules, optimizing for F1 score on TQ-IS. For a rule-based approach, we discover that optimal results can be achieved with only a small subset of the full ruleset. Using five rules, we obtain an F1 score of 98.2%. We then evaluate three unsupervised algorithms, i.e., Gaussian Mixture Models (GMMs), Isolation Forests and One-Class SVMs. Our findings reveal that unsupervised algorithms perform well on the TQ-IS dataset, with GMMs obtaining the best results, comparable to those obtained with the rule-based approach. Finally, we show that unsupervised methods appear to be equally suitable for languages other than Icelandic, including Estonian and Basque.

When parallel corpora are preprocessed for machine translation (MT) training, a part of the parallel data is commonly discarded and deemed non-parallel due to odd-length ratio, overlapping text in source and target sentences or failing some other form of a semantic equivalency test. For language pairs with limited parallel resources, this can be costly as in such cases modest amounts of acceptable data may be useful to help build MT systems that generate higher quality translations. In this paper, we refine parallel corpora for two language pairs, English–Bengali and English–Icelandic, by extracting sub-sentence fragments from sentence pairs that would otherwise have been discarded, in order to increase recall when compiling training data. We find that by including the fragments, translation quality of NMT systems trained on the data improves significantly when translating from English to Bengali and from English to Icelandic.

The methods used to create many of the well-known Question-Answering (QA) datasets are hard to replicate for low-resource languages. A commonality amongst these methods is hiring annotators to source answers from the internet by querying a single answer source, such as Wikipedia. Applying these methods for low-resource languages can be problematic since there is no single large answer source for these languages. Consequently, this can result in a high ratio of unanswered questions, since the amount of information in any single source is limited. To address this problem, we developed a novel crowd-sourcing platform to gather multiple-domain QA data for low-resource languages. Our platform, which consists of a mobile app and a web API, gamifies the data collection process. We successfully released the app for Icelandic (a low-resource language with about 350,000 native speakers) to build a dataset which rivals large QA datasets for high-resource languages both in terms of size and ratio of answered questions. We have made the platform open source with instructions on how to localize and deploy it to gather data for other low-resource languages.

We present SentAlign, an accurate sentence alignment tool designed to handle very large parallel document pairs. Given user-defined parameters, the alignment algorithm evaluates all possible alignment paths in fairly large documents of thousands of sentences and uses a divide-and-conquer approach to align documents containing tens of thousands of sentences. The scoring function is based on LaBSE bilingual sentence representations. SentAlign outperforms five other sentence alignment tools when evaluated on two different evaluation sets, German-French and English-Icelandic, and on a downstream machine translation task.

We train and evaluate four Part-of-Speech tagging models for Icelandic. Three are older models that obtained the highest accuracy for Icelandic when they were introduced. The fourth model is of a type that currently reaches state-of-the-art accuracy. We use the most recent version of the MIM-GOLD training/testing corpus, its newest tagset, and augmentation data to obtain results that are comparable between the various models. We examine the accuracy improvements with each model and analyse the errors produced by our transformer model, which is based on a previously published ConvBERT model. For the set of errors that all the models make, and for which they predict the same tag, we extract a random subset for manual inspection. Extrapolating from this subset, we obtain a lower bound estimate on annotation errors in the corpus as well as on some unsolvable tagging errors. We argue that further tagging accuracy gains for Icelandic can still be obtained by fixing the errors in MIM-GOLD and, furthermore, that it should still be possible to squeeze out some small gains from our transformer model.

This paper describes a collaborative European project whose aim was to gather open source Natural Language Processing (NLP) tools and make them accessible as running services and easy to try out in the European Language Grid (ELG). The motivation of the project was to increase accessibility for more European languages and make it easier for developers to use the underlying tools in their own applications. The project resulted in the containerization of 60 existing NLP tools for 16 languages, all of which are now currently running as easily testable services in the ELG platform.

We explore different approaches for filtering parallel data for MT training, whether the same filtering approaches suit different datasets, and if separate filters should be applied to a dataset depending on the translation direction. We evaluate the results of different approaches, both manually and on a downstream NMT task. We find that, first, it is beneficial to inspect how well different filtering approaches suit different datasets and, second, that while MT systems trained on data prepared using different filters do not differ substantially in quality, there is indeed a statistically significant difference. Finally, we find that the same training sets do not seem to suit different translation directions.

The work in progress on the CEF Action National Language Technology Platform (NLTP) is presented. The Action aims at combining the most advanced Language Technology (LT) tools and solutions in a new state-of-the-art, Artificial Intelli- gence (AI) driven, National Language Technology Platform (NLTP).

This article presents the work in progress on the collaborative project of several European countries to develop National Language Technology Platform (NLTP). The project aims at combining the most advanced Language Technology tools and solutions in a new, state-of-the-art, Artificial Intelligence driven, National Language Technology Platform for five EU/EEA official and lower-resourced languages.

Bilingual lexicons can be generated automatically using a wide variety of approaches. We perform a rigorous manual evaluation of four different methods: word alignments on different types of bilingual data, pivoting, machine translation and cross-lingual word embeddings. We investigate how the different setups perform using publicly available data for the English-Icelandic language pair, doing separate evaluations for each method, dataset and confidence class where it can be calculated. The results are validated by human experts, working with a random sample from all our experiments. By combining the most promising approaches and data sets, using confidence scores calculated from the data and the results of manually evaluating samples from our manual evaluation as indicators, we are able to induce lists of translations with a very high acceptance rate. We show how multiple different combinations generate lists with well over 90% acceptance rate, substantially exceeding the results for each individual approach, while still generating reasonably large candidate lists. All manually evaluated equivalence pairs are published in a new lexicon of over 232,000 pairs under an open license.

In this paper, we present the first Entity Linking corpus for Icelandic. We describe our approach of using a multilingual entity linking model (mGENRE) in combination with Wikipedia API Search (WAPIS) to label our data and compare it to an approach using WAPIS only. We find that our combined method reaches 53.9% coverage on our corpus, compared to 30.9% using only WAPIS. We analyze our results and explain the value of using a multilingual system when working with Icelandic. Additionally, we analyze the data that remain unlabeled, identify patterns and discuss why they may be more difficult to annotate.

In this paper, we evaluate several Transformer-based language models for Icelandic on four downstream tasks: Part-of-Speech tagging, Named Entity Recognition. Dependency Parsing, and Automatic Text Summarization. We pre-train four types of monolingual ELECTRA and ConvBERT models and compare our results to a previously trained monolingual RoBERTa model and the multilingual mBERT model. We find that the Transformer models obtain better results, often by a large margin, compared to previous state-of-the-art models. Furthermore, our results indicate that pre-training larger language models results in a significant reduction in error rates in comparison to smaller models. Finally, our results show that the monolingual models for Icelandic outperform a comparably sized multilingual model.

Being able to generate accurate word alignments is useful for a variety of tasks. While statistical word aligners can work well, especially when parallel training data are plentiful, multilingual embedding models have recently been shown to give good results in unsupervised scenarios. We evaluate an ensemble method for word alignment on four language pairs and demonstrate that by combining multiple tools, taking advantage of their different approaches, substantial gains can be made. This holds for settings ranging from very low-resource to high-resource. Furthermore, we introduce a new gold alignment test set for Icelandic and a new easy-to-use tool for creating manual word alignments.

Automatic Text Summarization (ATS) is the task of generating concise and fluent summaries from one or more documents. In this paper, we present IceSum, the first Icelandic corpus annotated with human-generated summaries. IceSum consists of 1,000 online news articles and their extractive summaries. We train and evaluate several neural network-based models on this dataset, comparing them against a selection of baseline methods. We find that an encoder-decoder model with a sequence-to-sequence based extractor obtains the best results, outperforming all baseline methods. Furthermore, we evaluate how the size of the training corpus affects the quality of the generated summaries. We release the corpus and the models with an open license.

Parallel sentences extracted from comparable corpora can be useful to supplement parallel corpora when training machine translation (MT) systems. This is even more prominent in low-resource scenarios, where parallel corpora are scarce. In this paper, we present a system which uses three very different measures to identify and score parallel sentences from comparable corpora. We measure the accuracy of our methods in low-resource settings by comparing the results against manually curated test data for English–Icelandic, and by evaluating an MT system trained on the concatenation of the parallel data extracted by our approach and an existing data set. We show that the system is capable of extracting useful parallel sentences with high accuracy, and that the extracted pairs substantially increase translation quality of an MT system trained on the data, as measured by automatic evaluation metrics.

In this paper, we describe a new national language technology programme for Icelandic. The programme, which spans a period of five years, aims at making Icelandic usable in communication and interactions in the digital world, by developing accessible, open-source language resources and software. The research and development work within the programme is carried out by a consortium of universities, institutions, and private companies, with a strong emphasis on cooperation between academia and industries. Five core projects will be the main content of the programme: language resources, speech recognition, speech synthesis, machine translation, and spell and grammar checking. We also describe other national language technology programmes and give an overview over the history of language technology in Iceland.

In this paper, we present a character-based BiLSTM model for splitting Icelandic compound words, and show how varying amounts of training data affects the performance of the model. Compounding is highly productive in Icelandic, and new compounds are constantly being created. This results in a large number of out-of-vocabulary (OOV) words, negatively impacting the performance of many NLP tools. Our model is trained on a dataset of 2.9 million unique word forms and their constituent structures from the Database of Icelandic Morphology. The model learns how to split compound words into two parts and can be used to derive the constituent structure of any word form. Knowing the constituent structure of a word form makes it possible to generate the optimal split for a given task, e.g., a full split for subword tokenization, or, in the case of part-of-speech tagging, splitting an OOV word until the largest known morphological head is found. The model outperforms other previously published methods when evaluated on a corpus of manually split word forms. This method has been integrated into Kvistur, an Icelandic compound word analyzer.

Parallel corpora are key to developing good machine translation systems. However, abundant parallel data are hard to come by, especially for languages with a low number of speakers. When rich morphology exacerbates the data sparsity problem, it is imperative to have accurate alignment and filtering methods that can help make the most of what is available by maximising the number of correctly translated segments in a corpus and minimising noise by removing incorrect translations and segments containing extraneous data. This paper sets out a research plan for improving alignment and filtering methods for parallel texts in low-resource settings. We propose an effective unsupervised alignment method to tackle the alignment problem. Moreover, we propose a strategy to supplement state-of-the-art models with automatically extracted information using basic NLP tools to effectively handle rich morphology.

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.

Previous work on using BiLSTM models for PoS tagging has primarily focused on small tagsets. We evaluate BiLSTM models for tagging Icelandic, a morphologically rich language, using a relatively large tagset. Our baseline BiLSTM model achieves higher accuracy than any other previously published tagger, when not taking advantage of a morphological lexicon. When we extend the model by incorporating such data, we outperform the earlier state-of-the-art results by a significant margin. We also report on work in progress that attempts to address the problem of data sparsity inherent to morphologically detailed, fine-grained tagsets. We experiment with training a separate model on only the lexical category and using the coarse-grained output tag as an input into to the main model. This method further increases the accuracy and reduces the tagging errors by 21.3% compared to previous state-of-the-art results. Finally, we train and test our tagger on a new gold standard for Icelandic.

We present an open-source, wide-coverage context-free grammar (CFG) for Icelandic, and an accompanying parsing system. The grammar has over 5,600 nonterminals, 4,600 terminals and 19,000 productions in fully expanded form, with feature agreement constraints for case, gender, number and person. The parsing system consists of an enhanced Earley-based parser and a mechanism to select best-scoring parse trees from shared packed parse forests. Our parsing system is able to parse about 90% of all sentences in articles published on the main Icelandic news websites. Preliminary evaluation with evalb shows an F-measure of 70.72% on parsed sentences. Our system demonstrates that parsing a morphologically rich language using a wide-coverage CFG can be practical.

In this paper, we describe the correction of PoS tags in a new Icelandic corpus, MIM-GOLD, consisting of about 1 million tokens sampled from the Tagged Icelandic Corpus, MÍM, released in 2013. The goal is to use the corpus, among other things, as a new gold standard for training and testing PoS taggers. The construction of the corpus was first described in 2010 together with preliminary work on error detection and correction. In this paper, we describe further the correction of tags in the corpus. We describe manual correction and a method for semi-automatic error detection and correction. We show that, even after manual correction, the number of tagging errors in the corpus can be reduced significantly by applying our semi-automatic detection and correction method. After the semi-automatic error correction, preliminary evaluation of tagging accuracy shows very low error rates. We hope that the existence of the corpus will make it possible to improve PoS taggers for Icelandic text.

This paper presents ongoing work that aims to improve machine parsing of Faroese using a combination of Faroese and Icelandic training data. We show that even if we only have a relatively small parsed corpus of one language, namely 53,000 words of Faroese, we can obtain better results by adding information about phrase structure from a closely related language which has a similar syntax. Our experiment uses the Berkeley parser. We demonstrate that the addition of Icelandic data without any other modification to the experimental setup results in an f-measure improvement from 75.44% to 78.05% in Faroese and an improvement in part-of-speech tagging accuracy from 88.86% to 90.40%.