Parallel corpora are still crucial to train effective Machine Translation systems. This is even more true for low-resource language pairs, for which Neural Machine Translation has been shown to be less robust to domain mismatch and noise. Due to time and resource constraints, parallel corpora are mostly created with sentence alignment methods which automatically infer alignments. Recent work focused on state-of-the-art pre-trained sentence embeddings-based methods which are available only for a tiny fraction of the world’s languages. In this paper, we evaluate the performance of four widely used algorithms on the low-resource English-Yorùbá language pair against a multidomain benchmark parallel corpus on two experiments involving 1-to-1 alignments with and without reordering. We find that, at least for this language pair, earlier and simpler methods are more suited to the task, all the while not requiring additional data or resources. We also report that the methods we evaluated perform differently across distinct domains, thus indicating that some approach may be better for a specific domain or textual structure.

The system is trained on officialy provided data only. We have heavily filtered all the data to remove machine translated text, Russian text and other noise. We use the DeepNorm modification of the transformer architecture in the TorchScale library with 18 encoder layers and 6 decoder layers. The initial systems for backtranslation uses HFT tokenizer, the final system uses custom tokenizer derived from HFT.

The WMT 2023 Shared Task on Low-Resource Indic Language Translation featured to and from Assamese, Khasi, Manipuri, Mizo on one side and English on the other. We submitted systems supervised neural machine translation systems for each pair and direction and experimented with different configurations and settings for both preprocessing and training. Even if most of them did not reach competitive performance, our experiments uncovered some interesting points for further investigation, namely the relation between dataset and model size, and the impact of the training framework. Moreover, the results of some of our preliminary experiments on the use of word embeddings initialization, backtranslation, and model depth were in contrast with previous work. The final results also show some disagreement in the automated metrics employed in the evaluation.

We describe our neural machine translation systems for the WMT22 shared task on unsupervised MT and very low resource supervised MT. We submit supervised NMT systems for Lower Sorbian-German and Lower Sorbian-Upper Sorbian translation in both directions. By using a novel tokenization algorithm, data augmentation techniques, such as Data Diversification (DD), and parameter optimization we improve on our baselines by 10.5-10.77 BLEU for Lower Sorbian-German and by 1.52-1.88 BLEU for Lower Sorbian-Upper Sorbian.

Tokenization has been shown to impact the quality of downstream tasks, such as Neural Machine Translation (NMT), which is susceptible to out-of-vocabulary words and low frequency training data. Current state-of-the-art algorithms have been helpful in addressing the issues of out-of-vocabulary words, bigger vocabulary sizes and token frequency by implementing subword segmentation. We argue, however, that there is still room for improvement, in particular regarding low-frequency tokens in the training data. In this paper, we present “High Frequency Tokenizer”, or HFT, a new language-independent subword segmentation algorithm that addresses this issue. We also propose a new metric to measure the frequency coverage of a tokenizer’s vocabulary, based on a frequency rank weighted average of the frequency values of its items. We experiment with a diverse set of language corpora, vocabulary sizes, and writing systems and report improvements on both frequency statistics and on the average length of the output. We also observe a positive impact on downstream NMT.

In this paper we discuss some of the current challenges in web corpus building that we faced in the recent years when expanding the corpora in Sketch Engine. The purpose of the paper is to provide an overview and raise discussion on possible solutions, rather than bringing ready solutions to the readers. For every issue we try to assess its severity and briefly discuss possible mitigation options.

In this paper we investigate two approaches to discrimination of similar languages: Expectation–maximization algorithm for estimating conditional probability P(word|language) and byte level language models similar to compression-based language modelling methods. The accuracy of these methods reached respectively 86.6% and 88.3% on set A of the DSL Shared task 2016 competition.

The paper describes automatic definition finding implemented within the leading corpus query and management tool, Sketch Engine. The implementation exploits complex pattern-matching queries in the corpus query language (CQL) and the indexing mechanism of word sketches for finding and storing definition candidates throughout the corpus. The approach is evaluated for Czech and English corpora, showing that the results are usable in practice: precision of the tool ranges between 30 and 75 percent (depending on the major corpus text types) and we were able to extract nearly 2 million definition candidates from an English corpus with 1.4 billion words. The feature is embedded into the interface as a concordance filter, so that users can search for definitions of any query to the corpus, including very specific multi-word queries. The results also indicate that ordinary texts (unlike explanatory texts) contain rather low number of definitions, which is perhaps the most important problem with automatic definition finding in general.

The NLP researcher or application-builder often wonders “what corpus should I use, or should I build one of my own? If I build one of my own, how will I know if I have done a good job?” Currently there is very little help available for them. They are in need of a framework for evaluating corpora. We develop such a framework, in relation to corpora which aim for good coverage of ‘general language’. The task we set is automatic creation of a publication-quality collocations dictionary. For a sample of 100 headwords of Czech and 100 of English, we identify a gold standard dataset of (ideally) all the collocations that should appear for these headwords in such a dictionary. The datasets are being made available alongside this paper. We then use them to determine precision and recall for a range of corpora, with a range of parameters.

We present HindEnCorp, a parallel corpus of Hindi and English, and HindMonoCorp, a monolingual corpus of Hindi in their release version 0.5. Both corpora were collected from web sources and preprocessed primarily for the training of statistical machine translation systems. HindEnCorp consists of 274k parallel sentences (3.9 million Hindi and 3.8 million English tokens). HindMonoCorp amounts to 787 million tokens in 44 million sentences. Both the corpora are freely available for non-commercial research and their preliminary release has been used by numerous participants of the WMT 2014 shared translation task.

In the paper the results of the project of Czech Legal Electronic dictionary (PES) are presented. During the 4 year project the large legal terminological dictionary of Czech was created in the form of the electronic lexical database enriched with a hierarchical ontology of legal terms. It contains approx. 10,000 entries ― legal terms together with their ontological relations and hypertext references. In the second part of the project the web interface based on the platform DEBII has been designed and implemented that allows users to browse and search effectively the database. At the same time the Czech Dictionary of Legal Terms will be generated from the database and later printed as a book. Inter-annotator's agreement in manual selection of legal terms was high ― approx. 95 %.

This work describes the process of creation of a 70 billion word text corpus of English. We used an existing language resource, namely the ClueWeb09 dataset, as source for the corpus data. Processing such a vast amount of data presented several challenges, mainly associated with pre-processing (boilerplate cleaning, text de-duplication) and post-processing (indexing for efficient corpus querying using the CQL -- Corpus Query Language) steps. In this paper we explain how we tackled them: we describe the tools used for boilerplate cleaning (jusText) and for de-duplication (onion) that was performed not only on full (document-level) duplicates but also on the level of near-duplicate texts. Moreover we show the impact of each of the performed pre-processing steps on the final corpus size. Furthermore we show how effective parallelization of the corpus indexation procedure was employed within the Manatee corpus management system and during computation of word sketches (one-page, automatic, corpus-derived summaries of a word's grammatical and collocational behaviour) from the resulting corpus.

We have analyzed the SPEX algorithm by Bernstein and Zobel (2004) for detecting co-derivative documents using duplicate n-grams. Although we totally agree with the claim that not using unique n-grams can greatly increase the efficiency and scalability of the process of detecting co-derivative documents, we have found serious bottlenecks in the way SPEX finds the duplicate n-grams. While the memory requirements for computing co-derivative documents can be reduced to up to 1% by only using duplicate n-grams, SPEX needs about 40 times more memory for computing the list of duplicate n-grams itself. Therefore the memory requirements of the whole process are not reduced enough to make the algorithm practical for very large collections. We propose a solution for this problem using an external sort with the suffix array in-memory sorting and temporary file compression. The proposed algorithm for computing duplicate n-grams uses a fixed amount of memory for any input size.