Large, curated, web-crawled corpora play a vital role in training language models (LMs). They form the lion’s share of the training data in virtually all recent LMs, such as the well-known GPT, LLaMA and XLM-RoBERTa models. However, despite this importance, relatively little attention has been given to the quality of these corpora. In this paper, we compare four of the currently most relevant large, web-crawled corpora (CC100, MaCoCu, mC4 and OSCAR) across eleven lower-resourced European languages. Our approach is two-fold: first, we perform an intrinsic evaluation by performing a human evaluation of the quality of samples taken from different corpora; then, we assess the practical impact of the qualitative differences by training specific LMs on each of the corpora and evaluating their performance on downstream tasks. We find that there are clear differences in quality of the corpora, with MaCoCu and OSCAR obtaining the best results. However, during the extrinsic evaluation, we actually find that the CC100 corpus achieves the highest scores. We conclude that, in our experiments, the quality of the web-crawled corpora does not seem to play a significant role when training LMs.

We present the most relevant results of the project MaCoCu: Massive collection and curation of monolingual and bilingual data: focus on under-resourced languages in its second year. To date, parallel and monolingual corpora have been produced for seven low-resourced European languages by crawling large amounts of textual data from selected top-level domains of the Internet; both human and automatic evaluation show its usefulness. In addition, several large language models pretrained on MaCoCu data have been published, as well as the code used to collect and curate the data.

Computer-aided translation (CAT) tools based on translation memories (MT) play a prominent role in the translation workflow of professional translators. However, the reduced availability of in-domain TMs, as compared to in-domain monolingual corpora, limits its adoption for a number of translation tasks. In this paper, we introduce a novel neural approach aimed at overcoming this limitation by exploiting not only TMs, but also in-domain target-language (TL) monolingual corpora, and still enabling a similar functionality to that offered by conventional TM-based CAT tools. Our approach relies on cross-lingual sentence embeddings to retrieve translation proposals from TL monolingual corpora, and on a neural model to estimate their post-editing effort. The paper presents an automatic evaluation of these techniques on four language pairs that shows that our approach can successfully exploit monolingual texts in a TM-based CAT environment, increasing the amount of useful translation proposals, and that our neural model for estimating the post-editing effort enables the combination of translation proposals obtained from monolingual corpora and from TMs in the usual way. A human evaluation performed on a single language pair confirms the results of the automatic evaluation and seems to indicate that the translation proposals retrieved with our approach are more useful than what the automatic evaluation shows.

We introduce the project “MaCoCu: Massive collection and curation of monolingual and bilingual data: focus on under-resourced languages”, funded by the Connecting Europe Facility, which is aimed at building monolingual and parallel corpora for under-resourced European languages. The approach followed consists of crawling large amounts of textual data from carefully selected top-level domains of the Internet, and then applying a curation and enrichment pipeline. In addition to corpora, the project will release successive versions of the free/open-source web crawling and curation software used.

An important goal of the MaCoCu project is to improve EU-specific NLP systems that concern their Digital Service Infrastructures (DSIs). In this paper we aim at boosting the creation of such domain-specific NLP systems. To do so, we explore the feasibility of building an automatic classifier that allows to identify which segments in a generic (potentially parallel) corpus are relevant for a particular DSI. We create an evaluation data set by crawling DSI-specific web domains and then compare different strategies to build our DSI classifier for text in three languages: English, Spanish and Dutch. We use pre-trained (multilingual) language models to perform the classification, with zero-shot classification for Spanish and Dutch. The results are promising, as we are able to classify DSIs with between 70 and 80% accuracy, even without in-language training data. A manual annotation of the data revealed that we can also find DSI-specific data on crawled texts from general web domains with reasonable accuracy. We publicly release all data, predictions and code, as to allow future investigations in whether exploiting this DSI-specific data actually leads to improved performance on particular applications, such as machine translation.

In the context of neural machine translation, data augmentation (DA) techniques may be used for generating additional training samples when the available parallel data are scarce. Many DA approaches aim at expanding the support of the empirical data distribution by generating new sentence pairs that contain infrequent words, thus making it closer to the true data distribution of parallel sentences. In this paper, we propose to follow a completely different approach and present a multi-task DA approach in which we generate new sentence pairs with transformations, such as reversing the order of the target sentence, which produce unfluent target sentences. During training, these augmented sentences are used as auxiliary tasks in a multi-task framework with the aim of providing new contexts where the target prefix is not informative enough to predict the next word. This strengthens the encoder and forces the decoder to pay more attention to the source representations of the encoder. Experiments carried out on six low-resource translation tasks show consistent improvements over the baseline and over DA methods aiming at extending the support of the empirical data distribution. The systems trained with our approach rely more on the source tokens, are more robust against domain shift and suffer less hallucinations.

In the media industry and the focus of global reporting can shift overnight. There is a compelling need to be able to develop new machine translation systems in a short period of time and in order to more efficiently cover quickly developing stories. As part of the EU project GoURMET and which focusses on low-resource machine translation and our media partners selected a surprise language for which a machine translation system had to be built and evaluated in two months(February and March 2021). The language selected was Pashto and an Indo-Iranian language spoken in Afghanistan and Pakistan and India. In this period we completed the full pipeline of development of a neural machine translation system: data crawling and cleaning and aligning and creating test sets and developing and testing models and and delivering them to the user partners. In this paperwe describe rapid data creation and experiments with transfer learning and pretraining for this low-resource language pair. We find that starting from an existing large model pre-trained on 50languages leads to far better BLEU scores than pretraining on one high-resource language pair with a smaller model. We also present human evaluation of our systems and which indicates that the resulting systems perform better than a freely available commercial system when translating from English into Pashto direction and and similarly when translating from Pashto into English.

We report on methods to create the largest publicly available parallel corpora by crawling the web, using open source software. We empirically compare alternative methods and publish benchmark data sets for sentence alignment and sentence pair filtering. We also describe the parallel corpora released and evaluate their quality and their usefulness to create machine translation systems.

This paper describes the joint submission of Universitat d’Alacant and Prompsit Language Engineering to the WMT 2020 shared task on parallel corpus filtering. Our submission, based on the free/open-source tool Bicleaner, enhances it with Extremely Randomised Trees and lexical similarity features that account for the frequency of the words in the parallel sentences to determine if two sentences are parallel. To train this classifier we used the clean corpora provided for the task and synthetic noisy parallel sentences. In addition we re-score the output of Bicleaner using character-level language models and n-gram saturation.

This paper describes our approach to create a neural machine translation system to translate between English and Swahili (both directions) in the news domain, as well as the process we followed to crawl the necessary parallel corpora from the Internet. We report the results of a pilot human evaluation performed by the news media organisations participating in the H2020 EU-funded project GoURMET.

Devising metrics to assess translation quality has always been at the core of machine translation (MT) research. Traditional automatic reference-based metrics, such as BLEU, have shown correlations with human judgements of adequacy and fluency and have been paramount for the advancement of MT system development. Crowd-sourcing has popularised and enabled the scalability of metrics based on human judgments, such as subjective direct assessments (DA) of adequacy, that are believed to be more reliable than reference-based automatic metrics. Finally, task-based measurements, such as post-editing time, are expected to provide a more de- tailed evaluation of the usefulness of translations for a specific task. Therefore, while DA averages adequacy judgements to obtain an appraisal of (perceived) quality independently of the task, and reference-based automatic metrics try to objectively estimate quality also in a task-independent way, task-based metrics are measurements obtained either during or after performing a specific task. In this paper we argue that, although expensive, task-based measurements are the most reliable when estimating MT quality in a specific task; in our case, this task is post-editing. To that end, we report experiments on a dataset with newly-collected post-editing indicators and show their usefulness when estimating post-editing effort. Our results show that task-based metrics comparing machine-translated and post-edited versions are the best at tracking post-editing effort, as expected. These metrics are followed by DA, and then by metrics comparing the machine-translated version and independent references. We suggest that MT practitioners should be aware of these differences and acknowledge their implications when decid- ing how to evaluate MT for post-editing purposes.

We describe the Universitat d’Alacant submissions to the word- and sentence-level machine translation (MT) quality estimation (QE) shared task at WMT 2018. Our approach to word-level MT QE builds on previous work to mark the words in the machine-translated sentence as OK or BAD, and is extended to determine if a word or sequence of words need to be inserted in the gap after each word. Our sentence-level submission simply uses the edit operations predicted by the word-level approach to approximate TER. The method presented ranked first in the sub-task of identifying insertions in gaps for three out of the six datasets, and second in the rest of them.

This paper presents an approach for building large monolingual corpora and, at the same time, extracting parallel data by crawling the top-level domain of a given language of interest. For gathering linguistically relevant data from top-level domains we use the SpiderLing crawler, modified to crawl data written in multiple languages. The output of this process is then fed to Bitextor, a tool for harvesting parallel data from a collection of documents. We call the system combining these two tools Spidextor, a blend of the names of its two crucial parts. We evaluate the described approach intrinsically by measuring the accuracy of the extracted bitexts from the Croatian top-level domain “.hr” and the Slovene top-level domain “.si”, and extrinsically on the English-Croatian language pair by comparing an SMT system built from the crawled data with third-party systems. We finally present parallel datasets collected with our approach for the English-Croatian, English-Finnish, English-Serbian and English-Slovene language pairs.

In this paper we compare two tools for automatically harvesting bitexts from multilingual websites: bitextor and ILSP-FC. We used both tools for crawling 21 multilingual websites from the tourism domain to build a domain-specific English―Croatian parallel corpus. Different settings were tried for both tools and 10,662 unique document pairs were obtained. A sample of about 10% of them was manually examined and the success rate was computed on the collection of pairs of documents detected by each setting. We compare the performance of the settings and the amount of different corpora detected by each setting. In addition, we describe the resource obtained, both by the settings and through the human evaluation, which has been released as a high-quality parallel corpus.

In this paper, a previous work on the enlargement of monolingual dictionaries of rule-based machine translation systems by non-expert users is extended to tackle the complete task of adding both source-language and target-language words to the monolingual dictionaries and the bilingual dictionary. In the original method, users validate whether some suffix variations of the word to be inserted are correct in order to find the most appropriate inflection paradigm. This method is now improved by taking advantage from the strong correlation detected between paradigms in both languages to reduce the search space of the target-language paradigm once the source-language paradigm is known. Results show that, when the source-language word has already been inserted, the system is able to more accurately predict which is the right target-language paradigm, and the number of queries posed to users is significantly reduced. Experiments also show that, when the source language and the target language are not closely related, it is only the source-language part-of-speech category, but not the rest of information provided by the source-language paradigm, which helps to correctly classify the target-language word.