While many new automatic metrics for machine translation evaluation have been proposed in recent years, BLEU scores are still used as the primary metric in the vast majority of MT research papers. There are many reasons that researchers may be reluctant to switch to new metrics, from external pressures (reviewers, prior work) to the ease of use of metric toolkits. Another reason is a lack of intuition about the meaning of novel metric scores. In this work, we examine “rules of thumb” about metric score differences and how they do (and do not) correspond to human judgments of statistically significant differences between systems. In particular, we show that common rules of thumb about BLEU score differences do not in fact guarantee that human annotators will find significant differences between systems. We also show ways in which these rules of thumb fail to generalize across translation directions or domains.

We argue that dialect identification should be treated as a multi-label classification problem rather than the single-class setting prevalent in existing collections and evaluations. In order to avoid extensive human re-labelling of the data, we propose an analysis of ambiguous near-duplicates in an existing collection covering four variants of French.We show how this analysis helps us provide multiple labels for a significant subset of the original data, therefore enriching the annotation with minimal human intervention. The resulting data can then be used to train dialect identifiers in a multi-label setting. Experimental results show that on the enriched dataset, the multi-label classifier produces similar accuracy to the single-label classifier on test cases that are unambiguous (single label), but it increases the macro-averaged F1-score by 0.225 absolute (71% relative gain) on ambiguous texts with multiple labels. On the original data, gains on the ambiguous test cases are smaller but still considerable (+0.077 absolute, 20% relative gain), and accuracy on non-ambiguous test cases is again similar in this case. This supports our thesis that modelling dialect identification as a multi-label problem potentially has a positive impact.

While recent studies have been dedicated to cleaning very noisy parallel corpora to improve Machine Translation training, we focus in this work on filtering a large and mostly clean Translation Memory. This problem of practical interest has not received much consideration from the community, in contrast with, for example, filtering large web-mined parallel corpora. We experiment with an extensive, multi-domain proprietary Translation Memory and compare five approaches involving deep-, feature-, and heuristic-based solutions. We propose two ways of evaluating this task, manual annotation and resulting Machine Translation quality. We report significant gains over a state-of-the-art, off-the-shelf cleaning system, using two MT engines.

We describe the systems developed by the National Research Council Canada for the French Cross-Domain Dialect Identification shared task at the 2022 VarDial evaluation campaign. We evaluated two different approaches to this task: SVM and probabilistic classifiers exploiting n-grams as features, and trained from scratch on the data provided; and a pre-trained French language model, CamemBERT, that we fine-tuned on the dialect identification task. The latter method turned out to improve the macro-F1 score on the test set from 0.344 to 0.430 (25% increase), which indicates that transfer learning can be helpful for dialect identification.

We describe the systems developed by the National Research Council Canada for the Uralic language identification shared task at the 2021 VarDial evaluation campaign. We evaluated two different approaches to this task: a probabilistic classifier exploiting only character 5-grams as features, and a character-based neural network pre-trained through self-supervision, then fine-tuned on the language identification task. The former method turned out to perform better, which casts doubt on the usefulness of deep learning methods for language identification, where they have yet to convincingly and consistently outperform simpler and less costly classification algorithms exploiting n-gram features.

We describe the systems developed by the National Research Council Canada for the Uralic language identification shared task at the 2020 VarDial evaluation campaign. Although our official results were well below the baseline, we show in this paper that this was not due to the neural approach to language identification in general, but to a flaw in the function we used to sample data for training and evaluation purposes. Preliminary experiments conducted after the evaluation period suggest that our neural approach to language identification can achieve state-of-the-art results on this task, although further experimentation is required.

Deep neural models tremendously improved machine translation. In this context, we investigate whether distinguishing machine from human translations is still feasible. We trained and applied 18 classifiers under two settings: a monolingual task, in which the classifier only looks at the translation; and a bilingual task, in which the source text is also taken into consideration. We report on extensive experiments involving 4 neural MT systems (Google Translate, DeepL, as well as two systems we trained) and varying the domain of texts. We show that the bilingual task is the easiest one and that transfer-based deep-learning classifiers perform best, with mean accuracies around 85% in-domain and 75% out-of-domain .

We describe the systems developed by the National Research Council Canada for the Cuneiform Language Identification (CLI) shared task at the 2019 VarDial evaluation campaign. We compare a state-of-the-art baseline relying on character n-grams and a traditional statistical classifier, a voting ensemble of classifiers, and a deep learning approach using a Transformer network. We describe how these systems were trained, and analyze the impact of some preprocessing and model estimation decisions. The deep neural network achieved 77% accuracy on the test data, which turned out to be the best performance at the CLI evaluation, establishing a new state-of-the-art for cuneiform language identification.

The WMT18 shared task on parallel corpus filtering (Koehn et al., 2018b) challenged teams to score sentence pairs from a large high-recall, low-precision web-scraped parallel corpus (Koehn et al., 2018a). Participants could use existing sample corpora (e.g. past WMT data) as a supervisory signal to learn what a “clean” corpus looks like. However, in lower-resource situations it often happens that the target corpus of the language is the only sample of parallel text in that language. We therefore made several unsupervised entries, setting ourselves an additional constraint that we not utilize the additional clean parallel corpora. One such entry fairly consistently scored in the top ten systems in the 100M-word conditions, and for one task—translating the European Medicines Agency corpus (Tiedemann, 2009)—scored among the best systems even in the 10M-word conditions.

We present our semantic textual similarity approach in filtering a noisy web crawled parallel corpus using YiSi—a novel semantic machine translation evaluation metric. The systems mainly based on this supervised approach perform well in the WMT18 Parallel Corpus Filtering shared task (4th place in 100-million-word evaluation, 8th place in 10-million-word evaluation, and 6th place overall, out of 48 submissions). In fact, our best performing system—NRC-yisi-bicov is one of the only four submissions ranked top 10 in both evaluations. Our submitted systems also include some initial filtering steps for scaling down the size of the test corpus and a final redundancy removal step for better semantic and token coverage of the filtered corpus. In this paper, we also describe our unsuccessful attempt in automatically synthesizing a noisy parallel development corpus for tuning the weights to combine different parallelism and fluency features.

Detecting changes within an unfolding event in real time from news articles or social media enables to react promptly to serious issues in public safety, public health or natural disasters. In this study, we use on-line Latent Dirichlet Allocation (LDA) to model shifts in topics, and apply on-line change point detection (CPD) algorithms to detect when significant changes happen. We describe an on-line Bayesian change point detection algorithm that we use to detect topic changes from on-line LDA output. Extensive experiments on social media data and news articles show the benefits of on-line LDA versus standard LDA, and of on-line change point detection compared to off-line algorithms. This yields F-scores up to 52% on the detection of significant real-life changes from these document streams.

Detecting events from social media data has important applications in public security, political issues, and public health. Many studies have focused on detecting specific or unspecific events from Twitter streams. However, not much attention has been paid to detecting changes, and their impact, in online conversations related to an event. We propose methods for detecting such changes, using clustering of temporal profiles of hashtags, and three change point detection algorithms. The methods were tested on two Twitter datasets: one covering the 2014 Ottawa shooting event, and one covering the Sochi winter Olympics. We compare our approach to a baseline consisting of detecting change from raw counts in the conversation. We show that our method produces large gains in change detection accuracy on both datasets.

We describe the submissions entered by the National Research Council Canada in the NLI-2017 evaluation. We mainly explored the use of voting, and various ways to optimize the choice and number of voting systems. We also explored the use of features that rely on no linguistic preprocessing. Long ngrams of characters obtained from raw text turned out to yield the best performance on all textual input (written essays and speech transcripts). Voting ensembles turned out to produce small performance gains, with little difference between the various optimization strategies we tried. Our top systems achieved accuracies of 87% on the essay track, 84% on the speech track, and close to 92% by combining essays, speech and i-vectors in the fusion track.

We describe the systems entered by the National Research Council in the 2016 shared task on discriminating similar languages. Like previous years, we relied on character ngram features, and a mixture of discriminative and generative statistical classifiers. We mostly investigated the influence of the amount of data on the performance, in the open task, and compared the two-stage approach (predicting language/group, then variant) to a flat approach. Results suggest that ngrams are still state-of-the-art for language and variant identification, and that additional data has a small but decisive impact.

We present an analysis of the performance of machine learning classifiers on discriminating between similar languages and language varieties. We carried out a number of experiments using the results of the two editions of the Discriminating between Similar Languages (DSL) shared task. We investigate the progress made between the two tasks, estimate an upper bound on possible performance using ensemble and oracle combination, and provide learning curves to help us understand which languages are more challenging. A number of difficult sentences are identified and investigated further with human annotation

The goal of keyphrase extraction is to automatically identify the most salient phrases from documents. The technique has a wide range of applications such as rendering a quick glimpse of a document, or extracting key content for further use. While previous work often assumes keyphrases are a static property of a given documents, in many applications, the appropriate set of keyphrases that should be extracted depends on the set of documents that are being considered together. In particular, good keyphrases should not only accurately describe the content of a document, but also reveal what discriminates it from the other documents. In this paper, we study this problem of extracting discriminative keyphrases. In particularly, we propose to use the hierarchical semantic structure between candidate keyphrases to promote keyphrases that have the right level of specificity to clearly distinguish the target document from others. We show that such knowledge can be used to construct better discriminative keyphrase extraction systems that do not assume a static, fixed set of keyphrases for a document. We show how this helps identify key expertise of authors from their papers, as well as competencies covered by online courses within different domains.

When parallel or comparable corpora are harvested from the web, there is typically a tradeoff between the size and quality of the data. In order to improve quality, corpus collection efforts often attempt to fix or remove misaligned sentence pairs. But, at the same time, Statistical Machine Translation (SMT) systems are widely assumed to be relatively robust to sentence alignment errors. However, there is little empirical evidence to support and characterize this robustness. This contribution investigates the impact of sentence alignment errors on a typical phrase-based SMT system. We confirm that SMT systems are highly tolerant to noise, and that performance only degrades seriously at very high noise levels. Our findings suggest that when collecting larger, noisy parallel data for training phrase-based SMT, cleaning up by trying to detect and remove incorrect alignments can actually degrade performance. Although fixing errors, when applicable, is a preferable strategy to removal, its benefits only become apparent for fairly high misalignment rates. We provide several explanations to support these findings.

Cet article présente une méthode de traduction automatique statistique basée sur des segments non-continus, c’est-à-dire des segments formés de mots qui ne se présentent pas nécéssairement de façon contiguë dans le texte. On propose une méthode pour produire de tels segments à partir de corpus alignés au niveau des mots. On présente également un modèle de traduction statistique capable de tenir compte de tels segments, de même qu’une méthode d’apprentissage des paramètres du modèle visant à maximiser l’exactitude des traductions produites, telle que mesurée avec la métrique NIST. Les traductions optimales sont produites par le biais d’une recherche en faisceau. On présente finalement des résultats expérimentaux, qui démontrent comment la méthode proposée permet une meilleure généralisation à partir des données d’entraînement.