We show that content on the web is often translated into many languages, and the low quality of these multi-way translations indicates they were likely created using Machine Translation (MT). Multi-way parallel, machine generated content not only dominates the translations in lower resource languages; it also constitutes a large fraction of the total web content in those languages. We also find evidence of a selection bias in the type of content which is translated into many languages, consistent with low quality English content being translated en masse into many lower resource languages, via MT. Our work raises serious concerns about training models such as multilingual large language models on both monolingual and bilingual data scraped from the web.

This paper reports on the shared tasks organized by the 21st IWSLT Conference. The shared tasks address 7 scientific challenges in spoken language translation: simultaneous and offline translation, automatic subtitling and dubbing, speech-to-speech translation, dialect and low-resource speech translation, and Indic languages. The shared tasks attracted 17 teams whose submissions are documented in 27 system papers. The growing interest towards spoken language translation is also witnessed by the constantly increasing number of shared task organizers and contributors to the overview paper, almost evenly distributed across industry and academia.

This paper reports on the shared tasks organized by the 20th IWSLT Conference. The shared tasks address 9 scientific challenges in spoken language translation: simultaneous and offline translation, automatic subtitling and dubbing, speech-to-speech translation, multilingual, dialect and low-resource speech translation, and formality control. The shared tasks attracted a total of 38 submissions by 31 teams. The growing interest towards spoken language translation is also witnessed by the constantly increasing number of shared task organizers and contributors to the overview paper, almost evenly distributed across industry and academia.

We explore zero-shot adaptation, where a general-domain model has access to customer or domain specific parallel data at inference time, but not during training. We build on the idea of Retrieval Augmented Translation (RAT) where top-k in-domain fuzzy matches are found for the source sentence, and target-language translations of those fuzzy-matched sentences are provided to the translation model at inference time. We propose a novel architecture to control interactions between a source sentence and the top-k fuzzy target-language matches, and compare it to architectures from prior work. We conduct experiments in two language pairs (En-De and En-Fr) by training models on WMT data and testing them with five and seven multi-domain datasets, respectively. Our approach consistently outperforms the alternative architectures, improving BLEU across language pair, domain, and number k of fuzzy matches.

Conventional speech-to-text translation (ST) systems are trained on single-speaker utterances, and they may not generalize to real-life scenarios where the audio contains conversations by multiple speakers. In this paper, we tackle single-channel multi-speaker conversational ST with an end-to-end and multi-task training model, named Speaker-Turn Aware Conversational Speech Translation, that combines automatic speech recognition, speech translation and speaker turn detection using special tokens in a serialized labeling format. We run experiments on the Fisher-CALLHOME corpus, which we adapted by merging the two single-speaker channels into one multi-speaker channel, thus representing the more realistic and challenging scenario with multi-speaker turns and cross-talk. Experimental results across single- and multi-speaker conditions and against conventional ST systems, show that our model outperforms the reference systems on the multi-speaker condition, while attaining comparable performance on the single-speaker condition. We release scripts for data processing and model training.

The machine translation (MT) task is typically formulated as that of returning a single translation for an input segment. However, in many cases, multiple different translations are valid and the appropriate translation may depend on the intended target audience, characteristics of the speaker, or even the relationship between speakers. Specific problems arise when dealing with honorifics, particularly translating from English into languages with formality markers. For example, the sentence “Are you sure?” can be translated in German as “Sind Sie sich sicher?” (formal register) or “Bist du dir sicher?” (informal). Using wrong or inconsistent tone may be perceived as inappropriate or jarring for users of certain cultures and demographics. This work addresses the problem of learning to control target language attributes, in this case formality, from a small amount of labeled contrastive data. We introduce an annotated dataset (CoCoA-MT) and an associated evaluation metric for training and evaluating formality-controlled MT models for six diverse target languages. We show that we can train formality-controlled models by fine-tuning on labeled contrastive data, achieving high accuracy (82% in-domain and 73% out-of-domain) while maintaining overall quality.

We present a very simple method for extending pretrained machine translation metrics to incorporate document-level context. We apply our method to four popular metrics: BERTScore, Prism, COMET, and the reference-free metric COMET-QE. We evaluate our document-level metrics on the MQM annotations from the WMT 2021 metrics shared task and find that the document-level metrics outperform their sentence-level counterparts in about 85% of the tested conditions, when excluding results on low-quality human references. Additionally, we show that our document-level extension of COMET-QE dramatically improves accuracy on discourse phenomena tasks, supporting our hypothesis that our document-level metrics are resolving ambiguities in the reference sentence by using additional context.

The evaluation campaign of the 19th International Conference on Spoken Language Translation featured eight shared tasks: (i) Simultaneous speech translation, (ii) Offline speech translation, (iii) Speech to speech translation, (iv) Low-resource speech translation, (v) Multilingual speech translation, (vi) Dialect speech translation, (vii) Formality control for speech translation, (viii) Isometric speech translation. A total of 27 teams participated in at least one of the shared tasks. This paper details, for each shared task, the purpose of the task, the data that were released, the evaluation metrics that were applied, the submissions that were received and the results that were achieved.

One key ingredient of neural machine translation is the use of large datasets from different domains and resources (e.g. Europarl, TED talks). These datasets contain documents translated by professional translators using different but consistent translation styles. Despite that, the model is usually trained in a way that neither explicitly captures the variety of translation styles present in the data nor translates new data in different and controllable styles. In this work, we investigate methods to augment the state of the art Transformer model with translator information that is available in part of the training data. We show that our style-augmented translation models are able to capture the style variations of translators and to generate translations with different styles on new data. Indeed, the generated variations differ significantly, up to +4.5 BLEU score difference. Despite that, human evaluation confirms that the translations are of the same quality.

The evaluation campaign of the International Conference on Spoken Language Translation (IWSLT 2021) featured this year four shared tasks: (i) Simultaneous speech translation, (ii) Offline speech translation, (iii) Multilingual speech translation, (iv) Low-resource speech translation. A total of 22 teams participated in at least one of the tasks. This paper describes each shared task, data and evaluation metrics, and reports results of the received submissions.

Sub-word segmentation is currently a standard tool for training neural machine translation (MT) systems and other NLP tasks. The goal is to split words (both in the source and target languages) into smaller units which then constitute the input and output vocabularies of the MT system. The aim of reducing the size of the input and output vocabularies is to increase the generalization capabilities of the translation model, enabling the system to translate and generate infrequent and new (unseen) words at inference time by combining previously seen sub-word units. Ideally, we would expect the created units to have some linguistic meaning, so that words are created in a compositional way. However, the most popular word-splitting method, Byte-Pair Encoding (BPE), which originates from the data compression literature, does not include explicit criteria to favor linguistic splittings nor to find the optimal sub-word granularity for the given training data. In this paper, we propose a statistically motivated extension of the BPE algorithm and an effective convergence criterion that avoids the costly experimentation cycle needed to select the best sub-word vocabulary size. Experimental results with morphologically rich languages show that our model achieves nearly-optimal BLEU scores and produces morphologically better word segmentations, which allows to outperform BPE’s generalization in the translation of sentences containing new words, as shown via human evaluation.

Language domains that require very careful use of terminology are abundant and reflect a significant part of the translation industry. In this work we introduce a benchmark for evaluating the quality and consistency of terminology translation, focusing on the medical (and COVID-19 specifically) domain for five language pairs: English to French, Chinese, Russian, and Korean, as well as Czech to German. We report the descriptions and results of the participating systems, commenting on the need for further research efforts towards both more adequate handling of terminologies as well as towards a proper formulation and evaluation of the task.

The COVID-19 pandemic is the worst pandemic to strike the world in over a century. Crucial to stemming the tide of the SARS-CoV-2 virus is communicating to vulnerable populations the means by which they can protect themselves. To this end, the collaborators forming the Translation Initiative for COvid-19 (TICO-19) have made test and development data available to AI and MT researchers in 35 different languages in order to foster the development of tools and resources for improving access to information about COVID-19 in these languages. In addition to 9 high-resourced, ”pivot” languages, the team is targeting 26 lesser resourced languages, in particular languages of Africa, South Asia and South-East Asia, whose populations may be the most vulnerable to the spread of the virus. The same data is translated into all of the languages represented, meaning that testing or development can be done for any pairing of languages in the set. Further, the team is converting the test and development data into translation memories (TMXs) that can be used by localizers from and to any of the languages.

The evaluation campaign of the International Conference on Spoken Language Translation (IWSLT 2020) featured this year six challenge tracks: (i) Simultaneous speech translation, (ii) Video speech translation, (iii) Offline speech translation, (iv) Conversational speech translation, (v) Open domain translation, and (vi) Non-native speech translation. A total of teams participated in at least one of the tracks. This paper introduces each track’s goal, data and evaluation metrics, and reports the results of the received submissions.

We present enhancements to a speech-to-speech translation pipeline in order to perform automatic dubbing. Our architecture features neural machine translation generating output of preferred length, prosodic alignment of the translation with the original speech segments, neural text-to-speech with fine tuning of the duration of each utterance, and, finally, audio rendering to enriches text-to-speech output with background noise and reverberation extracted from the original audio. We report and discuss results of a first subjective evaluation of automatic dubbing of excerpts of TED Talks from English into Italian, which measures the perceived naturalness of automatic dubbing and the relative importance of each proposed enhancement.

A variety of natural language tasks require processing of textual data which contains a mix of natural language and formal languages such as mathematical expressions. In this paper, we take unit conversions as an example and propose a data augmentation technique which lead to models learning both translation and conversion tasks as well as how to adequately switch between them for end-to-end localization.

This paper proposes a novel method to inject custom terminology into neural machine translation at run time. Previous works have mainly proposed modifications to the decoding algorithm in order to constrain the output to include run-time-provided target terms. While being effective, these constrained decoding methods add, however, significant computational overhead to the inference step, and, as we show in this paper, can be brittle when tested in realistic conditions. In this paper we approach the problem by training a neural MT system to learn how to use custom terminology when provided with the input. Comparative experiments show that our method is not only more effective than a state-of-the-art implementation of constrained decoding, but is also as fast as constraint-free decoding.

Neural Machine Translation (NMT) models generally perform translation using a fixed-size lexical vocabulary, which is an important bottleneck on their generalization capability and overall translation quality. The standard approach to overcome this limitation is to segment words into subword units, typically using some external tools with arbitrary heuristics, resulting in vocabulary units not optimized for the translation task. Recent studies have shown that the same approach can be extended to perform NMT directly at the level of characters, which can deliver translation accuracy on-par with subword-based models, on the other hand, this requires relatively deeper networks. In this paper, we propose a more computationally-efficient solution for character-level NMT which implements a hierarchical decoding architecture where translations are subsequently generated at the level of words and characters. We evaluate different methods for open-vocabulary NMT in the machine translation task from English into five languages with distinct morphological typology, and show that the hierarchical decoding model can reach higher translation accuracy than the subword-level NMT model using significantly fewer parameters, while demonstrating better capacity in learning longer-distance contextual and grammatical dependencies than the standard character-level NMT model.

The IWSLT 2019 evaluation campaign featured three tasks: speech translation of (i) TED talks and (ii) How2 instructional videos from English into German and Portuguese, and (iii) text translation of TED talks from English into Czech. For the first two tasks we encouraged submissions of end- to-end speech-to-text systems, and for the second task participants could also use the video as additional input. We received submissions by 12 research teams. This overview provides detailed descriptions of the data and evaluation conditions of each task and reports results of the participating systems.

Multilingual Neural Machine Translation (MNMT) for low- resource languages (LRL) can be enhanced by the presence of related high-resource languages (HRL), but the relatedness of HRL usually relies on predefined linguistic assumptions about language similarity. Recently, adapting MNMT to a LRL has shown to greatly improve performance. In this work, we explore the problem of adapting an MNMT model to an unseen LRL using data selection and model adapta- tion. In order to improve NMT for LRL, we employ perplexity to select HRL data that are most similar to the LRL on the basis of language distance. We extensively explore data selection in popular multilingual NMT settings, namely in (zero-shot) translation, and in adaptation from a multilingual pre-trained model, for both directions (LRL↔en). We further show that dynamic adaptation of the model’s vocabulary results in a more favourable segmentation for the LRL in comparison with direct adaptation. Experiments show re- ductions in training time and significant performance gains over LRL baselines, even with zero LRL data (+13.0 BLEU), up to +17.0 BLEU for pre-trained multilingual model dynamic adaptation with related data selection. Our method outperforms current approaches, such as massively multilingual models and data augmentation, on four LRL.

The recent advances introduced by neural machine translation (NMT) are rapidly expanding the application fields of machine translation, as well as reshaping the quality level to be targeted. In particular, if translations have to fit some given layout, quality should not only be measured in terms of adequacy and fluency, but also length. Exemplary cases are the translation of document files, subtitles, and scripts for dubbing, where the output length should ideally be as close as possible to the length of the input text. This pa-per addresses for the first time, to the best of our knowledge, the problem of controlling the output length in NMT. We investigate two methods for biasing the output length with a transformer architecture: i) conditioning the output to a given target-source length-ratio class and ii) enriching the transformer positional embedding with length information. Our experiments show that both methods can induce the network to generate shorter translations, as well as acquiring inter- pretable linguistic skills.

Neural machine translation models have shown to achieve high quality when trained and fed with well structured and punctuated input texts. Unfortunately, the latter condition is not met in spoken language translation, where the input is generated by an automatic speech recognition (ASR) system. In this paper, we study how to adapt a strong NMT system to make it robust to typical ASR errors. As in our application scenarios transcripts might be post-edited by human experts, we propose adaptation strategies to train a single system that can translate either clean or noisy input with no supervision on the input type. Our experimental results on a public speech translation data set show that adapting a model on a significant amount of parallel data including ASR transcripts is beneficial with test data of the same type, but produces a small degradation when translating clean text. Adapting on both clean and noisy variants of the same data leads to the best results on both input types.

The International Workshop of Spoken Language Translation (IWSLT) 2018 Evaluation Campaign featured two tasks: low-resource machine translation and speech translation. In the first task, manually transcribed speech had to be translated from Basque to English. Since this translation direction is a under-resourced language pair, participants were encouraged to use additional parallel data from related languages. In the second task, participants had to translate English audio into German text with a full speech-translation system. In the baseline condition, participants were free to use composite architectures, while in the end-to-end condition they were restricted to use a single model for the task. This year, eight research groups took part in the low-resource machine translation task and nine in the speech translation task.

We propose a method to transfer knowledge across neural machine translation (NMT) models by means of a shared dynamic vocabulary. Our approach allows to extend an initial model for a given language pair to cover new languages by adapting its vocabulary as long as new data become available (i.e., introducing new vocabulary items if they are not included in the initial model). The parameter transfer mechanism is evaluated in two scenarios: i) to adapt a trained single language NMT system to work with a new language pair and ii) to continuously add new language pairs to grow to a multilingual NMT system. In both the scenarios our goal is to improve the translation performance, while minimizing the training convergence time. Preliminary experiments spanning five languages with different training data sizes (i.e., 5k and 50k parallel sentences) show a significant performance gain ranging from +3.85 up to +13.63 BLEU in different language directions. Moreover, when compared with training an NMT model from scratch, our transfer-learning approach allows us to reach higher performance after training up to 4% of the total training steps.

In this paper we present an analysis of the two most prominent methodologies used for the human evaluation of MT quality, namely evaluation based on Post-Editing (PE) and evaluation based on Direct Assessment (DA). To this purpose, we exploit a publicly available large dataset containing both types of evaluations. We first focus on PE and investigate how sensitive TER-based evaluation is to the type and number of references used. Then, we carry out a comparative analysis of PE and DA to investigate the extent to which the evaluation results obtained by methodologies addressing different human perspectives are similar. This comparison sheds light not only on PE but also on the so-called reference bias related to monolingual DA. Also, we analyze if and how the two methodologies can complement each other’s weaknesses.

Multilingual neural machine translation (M-NMT) has recently shown to improve performance of machine translation of low-resource languages. Thanks to its implicit transfer-learning mechanism, the availability of a highly resourced language pair can be leveraged to learn useful representation for a lower resourced language. This work investigates how a low-resource translation task can be improved within a multilingual setting. First, we adapt a system trained on multiple language directions to a specific language pair. Then, we utilize the adapted model to apply an iterative training-inference scheme [1] using monolingual data. In the experimental setting, an extremely low-resourced Basque-English language pair (i.e., ≈ 5.6K in-domain training data) is our target translation task, where we considered a closely related French/Spanish-English parallel data to build the multilingual model. Experimental results from an i) in-domain and ii) an out-of-domain setting with additional training data, show improvements with our approach. We report a translation performance of 15.89 with the former and 23.99 BLEU with the latter on the official IWSLT 2018 Basque-English test set.

Neural machine translation (NMT) models are typically trained with fixed-size input and output vocabularies, which creates an important bottleneck on their accuracy and generalization capability. As a solution, various studies proposed segmenting words into sub-word units and performing translation at the sub-lexical level. However, statistical word segmentation methods have recently shown to be prone to morphological errors, which can lead to inaccurate translations. In this paper, we propose to overcome this problem by replacing the source-language embedding layer of NMT with a bi-directional recurrent neural network that generates compositional representations of the input at any desired level of granularity. We test our approach in a low-resource setting with five languages from different morphological typologies, and under different composition assumptions. By training NMT to compose word representations from character n-grams, our approach consistently outperforms (from 1.71 to 2.48 BLEU points) NMT learning embeddings of statistically generated sub-word units.

Recently, neural machine translation (NMT) has been extended to multilinguality, that is to handle more than one translation direction with a single system. Multilingual NMT showed competitive performance against pure bilingual systems. Notably, in low-resource settings, it proved to work effectively and efficiently, thanks to shared representation space that is forced across languages and induces a sort of transfer-learning. Furthermore, multilingual NMT enables so-called zero-shot inference across language pairs never seen at training time. Despite the increasing interest in this framework, an in-depth analysis of what a multilingual NMT model is capable of and what it is not is still missing. Motivated by this, our work (i) provides a quantitative and comparative analysis of the translations produced by bilingual, multilingual and zero-shot systems; (ii) investigates the translation quality of two of the currently dominant neural architectures in MT, which are the Recurrent and the Transformer ones; and (iii) quantitatively explores how the closeness between languages influences the zero-shot translation. Our analysis leverages multiple professional post-edits of automatic translations by several different systems and focuses both on automatic standard metrics (BLEU and TER) and on widely used error categories, which are lexical, morphology, and word order errors.

The requirement for neural machine translation (NMT) models to use fixed-size input and output vocabularies plays an important role for their accuracy and generalization capability. The conventional approach to cope with this limitation is performing translation based on a vocabulary of sub-word units that are predicted using statistical word segmentation methods. However, these methods have recently shown to be prone to morphological errors, which lead to inaccurate translations. In this paper, we extend the source-language embedding layer of the NMT model with a bi-directional recurrent neural network that generates compositional representations of the source words from embeddings of character n-grams. Our model consistently outperforms conventional NMT with sub-word units on four translation directions with varying degrees of morphological complexity and data sparseness on the source side.

Recurrent neural networks (RNNs) have represented for years the state of the art in neural machine translation. Recently, new architectures have been proposed, which can leverage parallel computation on GPUs better than classical RNNs. Faster training and inference combined with different sequence-to-sequence modeling also lead to performance improvements. While the new models completely depart from the original recurrent architecture, we decided to investigate how to make RNNs more efficient. In this work, we propose a new recurrent NMT architecture, called Simple Recurrent NMT, built on a class of fast and weakly-recurrent units that use layer normalization and multiple attentions. Our experiments on the WMT14 English-to-German and WMT16 English-Romanian benchmarks show that our model represents a valid alternative to LSTMs, as it can achieve better results at a significantly lower computational cost.

We address the issues arising when a neural machine translation engine trained on generic data receives requests from a new domain that contains many specific technical terms. Given training data of the new domain, we consider two alternative methods to adapt the generic system: corpus-based and instance-based adaptation. While the first approach is computationally more intensive in generating a domain-customized network, the latter operates more efficiently at translation time and can handle on-the-fly adaptation to multiple domains. Besides evaluating the generic and the adapted networks with conventional translation quality metrics, in this paper we focus on their ability to properly handle domain-specific terms. We show that instance-based adaptation, by fine-tuning the model on-the-fly, is capable to significantly boost the accuracy of translated terms, producing translations of quality comparable to the expensive corpusbased method.

This short presentation introduces ModernMT: an open-source project 1 that integrates real-time adaptive neural machine translation into a single easy-to-use product.

Both research and commercial machine translation have so far neglected the importance of properly handling the spelling, lexical and grammar divergences occurring among language varieties. Notable cases are standard national varieties such as Brazilian and European Portuguese, and Canadian and European French, which popular online machine translation services are not keeping distinct. We show that an evident side effect of modeling such varieties as unique classes is the generation of inconsistent translations. In this work, we investigate the problem of training neural machine translation from English to specific pairs of language varieties, assuming both labeled and unlabeled parallel texts, and low-resource conditions. We report experiments from English to two pairs of dialects, European-Brazilian Portuguese and European-Canadian French, and two pairs of standardized varieties, Croatian-Serbian and Indonesian-Malay. We show significant BLEU score improvements over baseline systems when translation into similar languages is learned as a multilingual task with shared representations.

State-of-the-art neural machine translation (NMT) systems are generally trained on specific domains by carefully selecting the training sets and applying proper domain adaptation techniques. In this paper we consider the real world scenario in which the target domain is not predefined, hence the system should be able to translate text from multiple domains. We compare the performance of a generic NMT system and phrase-based statistical machine translation (PBMT) system by training them on a generic parallel corpus composed of data from different domains. Our results on multi-domain English-French data show that, in these realistic conditions, PBMT outperforms its neural counterpart. This raises the question: is NMT ready for deployment as a generic/multi-purpose MT backbone in real-world settings?

The IWSLT 2017 evaluation campaign has organised three tasks. The Multilingual task, which is about training machine translation systems handling many-to-many language directions, including so-called zero-shot directions. The Dialogue task, which calls for the integration of context information in machine translation, in order to resolve anaphoric references that typically occur in human-human dialogue turns. And, finally, the Lecture task, which offers the challenge of automatically transcribing and translating real-life university lectures. Following the tradition of these reports, we will described all tasks in detail and present the results of all runs submitted by their participants.

Neural Machine Translation has been shown to enable inference and cross-lingual knowledge transfer across multiple language directions using a single multilingual model. Focusing on this multilingual translation scenario, this work summarizes FBK’s participation in the IWSLT 2017 shared task. Our submissions rely on two multilingual systems trained on five languages (English, Dutch, German, Italian, and Romanian). The first one is a 20 language direction model, which handles all possible combinations of the five languages. The second multilingual system is trained only on 16 directions, leaving the others as zero-shot translation directions (i.e representing a more complex inference task on language pairs not seen at training time). More specifically, our zero-shot directions are Dutch$German and Italian$Romanian (resulting in four language combinations). Despite the small amount of parallel data used for training these systems, the resulting multilingual models are effective, even in comparison with models trained separately for every language pair (i.e. in more favorable conditions). We compare and show the results of the two multilingual models against a baseline single language pair systems. Particularly, we focus on the four zero-shot directions and show how a multilingual model trained with small data can provide reasonable results. Furthermore, we investigate how pivoting (i.e using a bridge/pivot language for inference in a source!pivot!target translations) using a multilingual model can be an alternative to enable zero-shot translation in a low resource setting.

Neural machine translation (NMT) is the state of the art for machine translation, and it shows the best performance when there is a considerable amount of data available. When only little data exist for a language pair, the model cannot produce good representations for words, particularly for rare words. One common solution consists in reducing data sparsity by segmenting words into sub-words, in order to allow rare words to have shared representations with other words. Taking a different approach, in this paper we present a method to feed an NMT network with word embeddings trained on monolingual data, which are combined with the task-specific embeddings learned at training time. This method can leverage an embedding matrix with a huge number of words, which can therefore extend the word-level vocabulary. Our experiments on two language pairs show good results for the typical low-resourced data scenario (IWSLT in-domain dataset). Our consistent improvements over the baselines represent a positive proof about the possibility to leverage models pre-trained on monolingual data in NMT.

Recent work on multilingual neural machine translation reported competitive performance with respect to bilingual models and surprisingly good performance even on (zero-shot) translation directions not observed at training time. We investigate here a zero-shot translation in a particularly low-resource multilingual setting. We propose a simple iterative training procedure that leverages a duality of translations directly generated by the system for the zero-shot directions. The translations produced by the system (sub-optimal since they contain mixed language from the shared vocabulary), are then used together with the original parallel data to feed and iteratively re-train the multilingual network. Over time, this allows the system to learn from its own generated and increasingly better output. Our approach shows to be effective in improving the two zero-shot directions of our multilingual model. In particular, we observed gains of about 9 BLEU points over a baseline multilingual model and up to 2.08 BLEU over a pivoting mechanism using two bilingual models. Further analysis shows that there is also a slight improvement in the non-zero-shot language directions.

This paper presents WAGS (Word Alignment Gold Standard), a novel benchmark which allows extensive evaluation of WA tools on out-of-vocabulary (OOV) and rare words. WAGS is a subset of the Common Test section of the Europarl English-Italian parallel corpus, and is specifically tailored to OOV and rare words. WAGS is composed of 6,715 sentence pairs containing 11,958 occurrences of OOV and rare words up to frequency 15 in the Europarl Training set (5,080 English words and 6,878 Italian words), representing almost 3% of the whole text. Since WAGS is focused on OOV/rare words, manual alignments are provided for these words only, and not for the whole sentences. Two off-the-shelf word aligners have been evaluated on WAGS, and results have been compared to those obtained on an existing benchmark tailored to full text alignment. The results obtained confirm that WAGS is a valuable resource, which allows a statistically sound evaluation of WA systems’ performance on OOV and rare words, as well as extensive data analyses. WAGS is publicly released under a Creative Commons Attribution license.

The IWSLT 2016 Evaluation Campaign featured two tasks: the translation of talks and the translation of video conference conversations. While the first task extends previously offered tasks with talks from a different source, the second task is completely new. For both tasks, three tracks were organised: automatic speech recognition (ASR), spoken language translation (SLT), and machine translation (MT). Main translation directions that were offered are English to/from German and English to French. Additionally, the MT track included English to/from Arabic and Czech, as well as French to English. We received this year run submissions from 11 research labs. All runs were evaluated with objective metrics, while submissions for two of the MT talk tasks were also evaluated with human post-editing. Results of the human evaluation show improvements over the best submissions of last year.

In this paper, we describe FBK’s neural machine translation (NMT) systems submitted at the International Workshop on Spoken Language Translation (IWSLT) 2016. The systems are based on the state-of-the-art NMT architecture that is equipped with a bi-directional encoder and an attention mechanism in the decoder. They leverage linguistic information such as lemmas and part-of-speech tags of the source words in the form of additional factors along with the words. We compare performances of word and subword NMT systems along with different optimizers. Further, we explore different ensemble techniques to leverage multiple models within the same and across different networks. Several reranking methods are also explored. Our submissions cover all directions of the MSLT task, as well as en-{de, fr} and {de, fr}-en directions of TED. Compared to previously published best results on the TED 2014 test set, our models achieve comparable results on en-de and surpass them on en-fr (+2 BLEU) and fr-en (+7.7 BLEU) language pairs.

The paper overviews the 11th evaluation campaign organized by the IWSLT workshop. The 2014 evaluation offered multiple tracks on lecture transcription and translation based on the TED Talks corpus. In particular, this year IWSLT included three automatic speech recognition tracks, on English, German and Italian, five speech translation tracks, from English to French, English to German, German to English, English to Italian, and Italian to English, and five text translation track, also from English to French, English to German, German to English, English to Italian, and Italian to English. In addition to the official tracks, speech and text translation optional tracks were offered, globally involving 12 other languages: Arabic, Spanish, Portuguese (B), Hebrew, Chinese, Polish, Persian, Slovenian, Turkish, Dutch, Romanian, Russian. Overall, 21 teams participated in the evaluation, for a total of 76 primary runs submitted. Participants were also asked to submit runs on the 2013 test set (progress test set), in order to measure the progress of systems with respect to the previous year. All runs were evaluated with objective metrics, and submissions for two of the official text translation tracks were also evaluated with human post-editing.

This paper describes the systems submitted by FBK for the MT and SLT tracks of IWSLT 2014. We participated in the English-French and German-English machine translation tasks, as well as the English-French speech translation task. We report improvements in our English-French MT systems over last year’s baselines, largely due to improved techniques of combining translation and language models, and using huge language models. For our German-English system, we experimented with a novel domain adaptation technique. For both language pairs we also applied a novel word triggerbased model which shows slight improvements on EnglishFrench and German-English systems. Our English-French SLT system utilizes MT-based punctuation insertion, recasing, and ASR-like synthesized MT training data.

EU-BRIDGE is a European research project which is aimed at developing innovative speech translation technology. One of the collaborative efforts within EU-BRIDGE is to produce joint submissions of up to four different partners to the evaluation campaign at the 2014 International Workshop on Spoken Language Translation (IWSLT). We submitted combined translations to the German→English spoken language translation (SLT) track as well as to the German→English, English→German and English→French machine translation (MT) tracks. In this paper, we present the techniques which were applied by the different individual translation systems of RWTH Aachen University, the University of Edinburgh, Karlsruhe Institute of Technology, and Fondazione Bruno Kessler. We then show the combination approach developed at RWTH Aachen University which combined the individual systems. The consensus translations yield empirical gains of up to 2.3 points in BLEU and 1.2 points in TER compared to the best individual system.

In this paper we investigate the problem of adapting a machine translation system to the feedback provided by multiple post-editors. It is well know that translators might have very different post-editing styles and that this variability hinders the application of online learning methods, which indeed assume a homogeneous source of adaptation data. We hence propose multi-task learning to leverage bias information from each single post-editors in order to constrain the evolution of the SMT system. A new framework for significance testing with sentence level metrics is described which shows that Multi-Task learning approaches outperforms existing online learning approaches, with significant gains of 1.24 and 1.88 TER score over a strong online adaptive baseline, on a test set of post-edits produced by four translators texts and on a popular benchmark with multiple references, respectively.

Since the effectiveness of MT adaptation relies on the text repetitiveness, the question on how to measure repetitions in a text naturally arises. This work deals with the issue of looking for and evaluating text features that might help the prediction of the impact of MT adaptation on translation quality. In particular, the repetition rate metric, we recently proposed, is compared to other features employed in very related NLP tasks. The comparison is carried out through a regression analysis between feature values and MT performance gains by dynamically adapted versus non-adapted MT engines, on five different translation tasks. The main outcome of experiments is that the repetition rate correlates better than any other considered feature with the MT gains yielded by the online adaptation, although using all features jointly results in better predictions than with any single feature.

In spoken language translation, it is crucial that an automatic speech recognition (ASR) system produces outputs that can be adequately translated by a statistical machine translation (SMT) system. While word error rate (WER) is the standard metric of ASR quality, the assumption that each ASR error type is weighted equally is violated in a SMT system that relies on structured input. In this paper, we outline a statistical framework for analyzing the impact of specific ASR error types on translation quality in a speech translation pipeline. Our approach is based on linear mixed-effects models, which allow the analysis of ASR errors on a translation quality metric. The mixed-effects models take into account the variability of ASR systems and the difficulty of each speech utterance being translated in a specific experimental setting. We use mixed-effects models to verify that the ASR errors that compose the WER metric do not contribute equally to translation quality and that interactions exist between ASR errors that cumulatively affect a SMT system’s ability to translate an utterance. Our experiments are carried out on the English to French language pair using eight ASR systems and seven post-edited machine translation references from the IWSLT 2013 evaluation campaign. We report significant findings that demonstrate differences in the contributions of specific ASR error types toward speech translation quality and suggest further error types that may contribute to translation difficulty.

Defining the reordering search space is a crucial issue in phrase-based SMT between distant languages. In fact, the optimal trade-off between accuracy and complexity of decoding is nowadays reached by harshly limiting the input permutation space. We propose a method to dynamically shape such space and, thus, capture long-range word movements without hurting translation quality nor decoding time. The space defined by loose reordering constraints is dynamically pruned through a binary classifier that predicts whether a given input word should be translated right after another. The integration of this model into a phrase-based decoder improves a strong Arabic-English baseline already including state-of-the-art early distortion cost (Moore and Quirk, 2007) and hierarchical phrase orientation models (Galley and Manning, 2008). Significant improvements in the reordering of verbs are achieved by a system that is notably faster than the baseline, while bleu and meteor remain stable, or even increase, at a very high distortion limit.

The paper overviews the tenth evaluation campaign organized by the IWSLT workshop. The 2013 evaluation offered multiple tracks on lecture transcription and translation based on the TED Talks corpus. In particular, this year IWSLT included two automatic speech recognition tracks, on English and German, three speech translation tracks, from English to French, English to German, and German to English, and three text translation track, also from English to French, English to German, and German to English. In addition to the official tracks, speech and text translation optional tracks were offered involving 12 other languages: Arabic, Spanish, Portuguese (B), Italian, Chinese, Polish, Persian, Slovenian, Turkish, Dutch, Romanian, Russian. Overall, 18 teams participated in the evaluation for a total of 217 primary runs submitted. All runs were evaluated with objective metrics on a current test set and two progress test sets, in order to compare the progresses against systems of the previous years. In addition, submissions of one of the official machine translation tracks were also evaluated with human post-editing.

EU-BRIDGE1 is a European research project which is aimed at developing innovative speech translation technology. This paper describes one of the collaborative efforts within EUBRIDGE to further advance the state of the art in machine translation between two European language pairs, English→French and German→English. Four research institutions involved in the EU-BRIDGE project combined their individual machine translation systems and participated with a joint setup in the machine translation track of the evaluation campaign at the 2013 International Workshop on Spoken Language Translation (IWSLT). We present the methods and techniques to achieve high translation quality for text translation of talks which are applied at RWTH Aachen University, the University of Edinburgh, Karlsruhe Institute of Technology, and Fondazione Bruno Kessler. We then show how we have been able to considerably boost translation performance (as measured in terms of the metrics BLEU and TER) by means of system combination. The joint setups yield empirical gains of up to 1.4 points in BLEU and 2.8 points in TER on the IWSLT test sets compared to the best single systems.

This paper describes the systems submitted by FBK for the MT track of IWSLT 2013. We participated in the English-French as well as the bidirectional Persian-English translation tasks. We report substantial improvements in our English-French systems over last year’s baselines, largely due to improved techniques of combining translation and language models. For our Persian-English and English-Persian systems, we observe substantive improvements over baselines submitted by the workshop organizers, due to enhanced language-specific text normalization and the creation of a large monolingual news corpus in Persian.

This paper provides a fast alternative to Minimum Discrimination Information-based language model adaptation for statistical machine translation. We provide an alternative to computing a normalization term that requires computing full model probabilities (including back-off probabilities) for all n-grams. Rather than re-estimating an entire language model, our Lazy MDI approach leverages a smoothed unigram ratio between an adaptation text and the background language model to scale only the n-gram probabilities corresponding to translation options gathered by the SMT decoder. The effects of the unigram ratio are scaled by adding an additional feature weight to the log-linear discriminative model. We present results on the IWSLT 2012 TED talk translation task and show that Lazy MDI provides comparable language model adaptation performance to classic MDI.

We report here on the eighth evaluation campaign organized in 2011 by the IWSLT workshop series. That IWSLT 2011 evaluation focused on the automatic translation of public talks and included tracks for speech recognition, speech translation, text translation, and system combination. Unlike in previous years, all data supplied for the evaluation has been publicly released on the workshop website, and is at the disposal of researchers interested in working on our benchmarks and in comparing their results with those published at the workshop. This paper provides an overview of the IWSLT 2011 evaluation campaign, and describes the data supplied, the evaluation infrastructure made available to participants, and the subjective evaluation carried out.

This paper addresses the problem of reliably measuring productivity gains by professional translators working with a machine translation enhanced computer assisted translation tool. In particular, we report on a field test we carried out with a commercial CAT tool in which translation memory matches were supplemented with suggestions from a commercial machine translation engine. The field test was conducted with 12 professional translators working on real translation projects. Productivity of translators were measured with two indicators, post-editing speed and post-editing effort, on two translation directions, English–Italian and English–German, and two linguistic domains, legal and information technology. Besides a detailed statistical analysis of the experimental results, we also discuss issues encountered in running the test.

Several studies have recently reported significant productivity gains by human translators when besides translation memory (TM) matches they do also receive suggestions from a statistical machine translation (SMT) engine. In fact, an increasing number of language service providers and in-house translation services of large companies is nowadays integrating SMT in their workflow. The technology transfer of state-of-the-art SMT technology from research to industry has been relatively fast and simple also thanks to development of open source software, such as MOSES, GIZA++, and IRSTLM. While a translator is working on a specific translation project, she evaluates the utility of translating versus post-editing a segment based on the adequacy and fluency provided by the SMT engine, which in turn depends on the considered language pair, linguistic domain of the task, and the amount of available training data. Statistical models, like those employed in SMT, rely on a simple assumption: data used to train and tune the models represent the target translation task. Unfortunately, this assumption cannot be satisfied for most of the real application cases, simply because for most of the language pairs and domains there is no sufficient data to adequately train an SMT system. Hence, common practice is to train SMT systems by merging together parallel and monolingual data from the target domain with as much as possible data from any other available source. This workaround is simple and gives practical benefits but is often not the best way to exploit the available data. This tutorial copes with the optimal use of in-domain and out-of-domain data to achieve better SMT performance on a given application domain. Domain adaptation, in general, refers to statistical modeling and machine learning techniques that try to cope with the unavoidable mismatch between training and task data that typically occurs in real life applications. Our tutorial will survey several application cases in which domain adaptation can be applied, and presents adaptation techniques that best fit each case. In particular, we will cover adaptation methods for n-gram language models and translation models in phrase-based SMT. The tutorial will provide some high-level theoretical background in domain adaptation, it will discuss practical application cases, and finally show how the presented methods can be applied with two widely used software tools: Moses and IRSTLM. The tutorial is suited for any practitioner of statistical machine translation. No particular programming or mathematical background is required.

We report here on the eighth Evaluation Campaign organized by the IWSLT workshop. This year, the IWSLT evaluation focused on the automatic translation of public talks and included tracks for speech recognition, speech translation, text translation, and system combination. Unlike previous years, all data supplied for the evaluation has been publicly released on the workshop website, and is at the disposal of researchers interested in working on our benchmarks and in comparing their results with those published at the workshop. This paper provides an overview of the IWSLT 2011 Evaluation Campaign, which includes: descriptions of the supplied data and evaluation specifications of each track, the list of participants specifying their submitted runs, a detailed description of the subjective evaluation carried out, the main findings of each exercise drawn from the results and the system descriptions prepared by the participants, and, finally, several detailed tables reporting all the evaluation results.

This paper compares techniques to combine diverse parallel corpora for domain-specific phrase-based SMT system training. We address a common scenario where little in-domain data is available for the task, but where large background models exist for the same language pair. In particular, we focus on phrase table fill-up: a method that effectively exploits background knowledge to improve model coverage, while preserving the more reliable information coming from the in-domain corpus. We present experiments on an emerging transcribed speech translation task – the TED talks. While performing similarly in terms of BLEU and NIST scores to the popular log-linear and linear interpolation techniques, filled-up translation models are more compact and easy to tune by minimum error training.

Many syntactic machine translation decoders, including Moses, cdec, and Joshua, implement bottom-up dynamic programming to integrate N-gram language model probabilities into hypothesis scoring. These decoders concatenate hypotheses according to grammar rules, yielding larger hypotheses and eventually complete translations. When hypotheses are concatenated, the language model score is adjusted to account for boundary-crossing n-grams. Words on the boundary of each hypothesis are encoded in state, consisting of left state (the first few words) and right state (the last few words). We speed concatenation by encoding left state using data structure pointers in lieu of vocabulary indices and by avoiding unnecessary queries. To increase the decoder’s opportunities to recombine hypothesis, we minimize the number of words encoded by left state. This has the effect of reducing search errors made by the decoder. The resulting gain in model score is smaller than for right state minimization, which we explain by observing a relationship between state minimization and language model probability. With a fixed cube pruning pop limit, we show a 3-6% reduction in CPU time and improved model scores. Reducing the pop limit to the point where model scores tie the baseline yields a net 11% reduction in CPU time.

This paper gives an overview of the evaluation campaign results of the 7th International Workshop on Spoken Language Translation (IWSLT 2010)1. This year, we focused on three spoken language tasks: (1) public speeches on a variety of topics (TALK) from English to French, (2) spoken dialog in travel situations (DIALOG) between Chinese and English, and (3) traveling expressions (BTEC) from Arabic, Turkish, and French to English. In total, 28 teams (including 7 firsttime participants) took part in the shared tasks, submitting 60 primary and 112 contrastive runs. Automatic and subjective evaluations of the primary runs were carried out in order to investigate the impact of different communication modalities, spoken language styles and semantic context on automatic speech recognition (ASR) and machine translation (MT) system performances.

This year FBK took part in the BTEC translation task, with source languages Arabic and Turkish and target language English, and in the new TALK task, source English and target French. We worked in the framework of phrase-based statistical machine translation aiming to improve coverage of models in presence of rich morphology, on one side, and to make better use of available resources through data selection techniques. New morphological segmentation rules were developed for Turkish-English. The combination of several Turkish segmentation schemes into a lattice input led to an improvement wrt to last year. The use of additional training data was explored for Arabic-English, while on the English to French task improvement was achieved over a strong baseline by automatically selecting relevant and high quality data from the available training corpora.

This paper proposes a novel method for exploiting comparable documents to generate parallel data for machine translation. First, each source document is paired to each sentence of the corresponding target document; second, partial phrase alignments are computed within the paired texts; finally, fragment pairs across linked phrase-pairs are extracted. The algorithm has been tested on two recent challenging news translation tasks. Results show that mining for parallel fragments is more effective than mining for parallel sentences, and that comparable in-domain texts can be more valuable than parallel out-of-domain texts.

Current Statistical Machine Translation (SMT) systems translate texts sentence by sentence without considering any cross-sentential context. Assuming independence between sentences makes it difficult to take certain translation decisions when the necessary information cannot be determined locally. We argue for the necessity to include crosssentence dependencies in SMT. As a case in point, we study the problem of pronominal anaphora translation by manually evaluating German-English SMT output. We then present a word dependency model for SMT, which can represent links between word pairs in the same or in different sentences. We use this model to integrate the output of a coreference resolution system into English-German SMT with a view to improving the translation of anaphoric pronouns.

This paper reports on the participation of FBK at the IWSLT 2009 Evaluation. This year we worked on the Arabic-English and Turkish-English BTEC tasks with a special effort on linguistic preprocessing techniques involving morphological segmentation. In addition, we investigated the adaptation problem in the development of systems for the Chinese-English and English-Chinese challenge tasks; in particular, we explored different ways for clustering training data into topic or dialog-specific subsets: by producing (and combining) smaller but more focused models, we intended to make better use of the available training data, with the ultimate purpose of improving translation quality.

We tried to cope with the complex morphology of Turkish by applying different schemes of morphological word segmentation to the training and test data of a phrase-based statistical machine translation system. These techniques allow for a considerable reduction of the training dictionary, and lower the out-of-vocabulary rate of the test set. By minimizing differences between lexical granularities of Turkish and English we can produce more refined alignments and a better modeling of the translation task. Morphological segmentation is highly language dependent and requires a fair amount of linguistic knowledge in its development phase. Yet it is fast and light-weight – does not involve syntax – and appears to benefit our IWSLT09 system: our best segmentation scheme associated to a simple lexical approximation technique achieved a 50% reduction of out-of-vocabulary rate and over 5 point BLEU improvement above the baseline.

This paper focuses on the problem of language model adaptation in the context of Chinese-English cross-lingual dialogs, as set-up by the challenge task of the IWSLT 2009 Evaluation Campaign. Mixtures of n-gram language models are investigated, which are obtained by clustering bilingual training data according to different available human annotations, respectively, at the dialog level, turn level, and dialog act level. For the latter case, clustering of IWSLT data was in fact induced through a comparable Italian-English parallel corpus provided with dialog act annotations. For the sake of adaptation, mixture weight estimation is performed either at the level of single source sentence or test set. Estimated weights are then transferred to the target language mixture model. Experimental results show that, by training different specific language models weighted according to the actual input instead of using a single target language model, significant gains in terms of perplexity and BLEU can be achieved.

This paper reports on the participation of FBK at the IWSLT 2008 Evaluation. Main effort has been spent on the Chinese-Spanish Pivot task. We implemented four methods to perform pivot translation. The results on the IWSLT 2008 test data show that our original method for generating training data through random sampling outperforms the best methods based on coupling translation systems. FBK also participated in the Chinese-English Challenge task and the Chinese-English and Chinese-Spanish BTEC tasks, employing the standard state-of-the-art MT system Moses Toolkit.

Translation with pivot languages has recently gained attention as a means to circumvent the data bottleneck of statistical machine translation (SMT). This paper tries to give a mathematically sound formulation of the various approaches presented in the literature and introduces new methods for training alignment models through pivot languages. We present experimental results on Chinese-Spanish translation via English, on a popular traveling domain task. In contrast to previous literature, we report experimental results by using parallel corpora that are either disjoint or overlapped on the pivot language side. Finally, our original method for generating training data through random sampling shows to perform as well as the best methods based on the coupling of translation systems.

This work extends phrase-based statistical MT (SMT) with shallow syntax dependencies. Two string-to-chunks translation models are proposed: a factored model, which augments phrase-based SMT with layered dependencies, and a joint model, that extends the phrase translation table with microtags, i.e. per-word projections of chunk labels. Both rely on n-gram models of target sequences with different granularity: single words, micro-tags, chunks. In particular, n-grams defined over syntactic chunks should model syntactic constraints coping with word-group movements. Experimental analysis and evaluation conducted on two popular Chinese-English tasks suggest that the shallow-syntax joint-translation model has potential to outperform state-of-the-art phrase-based translation, with a reasonable computational overhead.

This paper reports on the participation of FBK (formerly ITC-irst) at the IWSLT 2007 Evaluation. FBK participated in three tasks, namely Chinese-to-English, Japanese-to-English, and Italian-to-English. With respect to last year, translation systems were developed with the Moses Toolkit and the IRSTLM library, both available as open source software. Moreover, several novel ideas were investigated: the use of confusion networks in input to manage ambiguity in punctuation, the estimation of an additional language model by means of the Google’s Web 1T 5-gram collection, the combination of true case and lower case language models, and finally the use of multiple phrase-tables. By working on top of a state-of-the art baseline, experiments showed that the above methods accounted for significant BLEU score improvements.