Multilingual pre-trained language models are often the best alternative in low-resource settings. In the context of a cascade architecture for automatic Standard German captioning of spoken Swiss German, we evaluate different models on the task of transforming normalised Swiss German ASR output into Standard German. Instead of training a large model from scratch, we fine-tuned publicly available pre-trained models, which reduces the cost of training high-quality neural machine translation models. Results show that pre-trained multilingual models achieve the highest scores, and that a higher number of languages included in pre-training improves the performance. We also observed that the type of source and target included in fine-tuning data impacts the results.

PROPICTO is a project funded by the French National Research Agency and the Swiss National Science Foundation, that aims at creating Speech-to-Pictograph translation systems, with a special focus on French as an input language. By developing such technologies, we intend to enhance communication access for non-French speaking patients and people with cognitive impairments.

Machine translation, and more specifically neural machine translation (NMT), have been proven to be subject to gender bias in recent years. Many studies have focused on evaluating and reducing this phenomenon, mainly through the analysis of occupational nouns’ translation for the same type of language combinations. In this paper, we reproduce a similar test set than in previous studies to investigate the influence of stereotypes and language combinations’ nature (formed with English, French and Italian) on gender bias occurrence in NMT. Similarly to previous studies, we confirm stereotypes as a major source of gender bias, especially in female contexts, while observing bias even in language combinations traditionally less examined.

The past few years have seen the multiplication of studies on post-editese, following the massive adoption of post-editing in professional translation workflows. These studies mainly rely on the comparison of post-edited machine translation and human translation on artificial parallel corpora. By contrast, we investigate here post-editese on comparable corpora of authentic translation jobs for the language direction English into French. We explore commonly used scores and also proposes the use of a novel metric. Our analysis shows that post-edited machine translation is not only lexically poorer than human translation, but also less dense and less varied in terms of translation solutions. It also tends to be more prolific than human translation for our language direction. Finally, our study highlights some of the challenges of working with comparable corpora in post-editese research.

We present the PASSAGE project, which aims at automatic Standard German subtitling of Swiss German TV content. This is achieved in a two step process, beginning with ASR to produce a normalised transcription, followed by translation into Standard German. We focus on the second step, for which we explore different approaches and contribute aligned corpora for future research.

The use of images has been shown to positively affect patient comprehension in medical settings, in particular to deliver specific medical instructions. However, tools that automatically translate sentences into pictographs are still scarce due to the lack of resources. Previous studies have focused on the translation of sentences into pictographs by using WordNet combined with rule-based approaches and deep learning methods. In this work, we showed how we leveraged the BabelDr system, a speech to speech translator for medical triage, to build a speech to pictograph translator using UMLS and neural machine translation approaches. We showed that the translation from French sentences to a UMLS gloss can be viewed as a machine translation task and that a Multilingual Neural Machine Translation system achieved the best results.

In Switzerland, two thirds of the population speak Swiss German, a primarily spoken language with no standardised written form. It is widely used on Swiss TV, for example in news reports, interviews or talk shows, and subtitles are required for people who cannot understand this spoken language. This paper focuses on the task of automatic Standard German subtitling of spoken Swiss German, and more specifically on the translation of a normalised Swiss German speech recognition result into Standard German suitable for subtitles. Our contribution consists of a comparison of different statistical and deep learning MT systems for this task and an aligned corpus of normalised Swiss German and Standard German subtitles. Results of two evaluations, automatic and human, show that the systems succeed in improving the content, but are currently not capable of producing entirely correct Standard German.

In this overview article we describe an application designed to enable communication between health practitioners and patients who do not share a common language, in situations where professional interpreters are not available. Built on the principle of a fixed phrase translator, the application implements different natural language processing (NLP) technologies, such as speech recognition, neural machine translation and text-to-speech to improve usability. Its design allows easy portability to new domains and integration of different types of output for multiple target audiences. Even though BabelDr is far from solving the problem of miscommunication between patients and doctors, it is a clear example of NLP in a real world application designed to help minority groups to communicate in a medical context. It also gives some insights into the relevant criteria for the development of such an application.

In international organizations, the growing demand for translations has increased the need for post-editing. Different studies show that automatic speech recognition systems have the potential to increase the productivity of the translation process as well as the quality. In this talk, we will explore the possibilities of using speech in the translation process by conducting a post-editing experiment with three professional translators in an international organization. Our experiment consisted of comparing three translation methods: speaking the translation with MT as an inspiration (RESpeaking), post-editing the MT suggestions by typing (PE), and editing the MT suggestion using speech (SPE). BLEU and HTER scores were used to compare the three methods. Our study shows that translators did more edits under condition RES, whereas in SPE, the resulting translations were closer to the reference according to the BLEU score and required less edits. Time taken to translate was the least in SPE followed by PE, RES methods and the translators preferred using speech to typing. These results show the potential of speech when it is coupled with post-editing. To the best of our knowledge, this is the first quantitative study conducted on using post-editing and speech together in large scale international organizations.

We present an adaptation of the Text-to-Picto system, initially designed for Dutch, and extended to English and Spanish. The original system, aimed at people with an intellectual disability, automatically translates text into pictographs (Sclera and Beta). We extend it to French and add a large set of Arasaac pictographs linked to WordNet 3.1. To carry out this adaptation, we automatically link the pictographs and their metadata to synsets of two French WordNets and leverage this information to translate words into pictographs. We automatically and manually evaluate our system with different corpora corresponding to different use cases, including one for medical communication between doctors and patients. The system is also compared to similar systems in other languages.

In diagnostic interviews, elliptical utterances allow doctors to question patients in a more efficient and economical way. However, literal translation of such incomplete utterances is rarely possible without affecting communication. Previous studies have focused on automatic ellipsis detection and resolution, but only few specifically address the problem of automatic translation of ellipsis. In this work, we evaluate four different approaches to translate ellipsis in medical dialogues in the context of the speech to speech translation system BabelDr. We also investigate the impact of training data, using an under-sampling method and data with elliptical utterances in context. Results show that the best model is able to translate 88% of elliptical utterances.

We believe that machine translation (MT) must be introduced to translation students as part of their training, in preparation for their professional life. In this paper we present a new version of the tool called MT3, which builds on and extends a joint effort undertaken by the Faculty of Languages of the University of Córdoba and Faculty of Translation and Interpreting of the University of Geneva to develop an open-source web platform to teach MT to translation students. We also report on a pilot experiment with the goal of testing the viability of using MT³ in an MT course. The pilot let us identify areas for improvement and collect students’ feedback about the tool’s usability.

In this study, we compare the output quality of two MT systems, a statistical (SMT) and a neural (NMT) engine, customised for Swiss Post’s Language Service using the same training data. We focus on the point of view of professional translators and investigate how they perceive the differences between the MT output and a human reference (namely deletions, substitutions, insertions and word order). Our findings show that translators more frequently consider these differences to be errors in SMT than NMT, and that deletions are the most serious errors in both architectures. We also observe lower agreement on differences to be corrected in NMT than in SMT, suggesting that errors are easier to identify in SMT. These findings confirm the ability of NMT to produce correct paraphrases, which could also explain why BLEU is often considered as an inadequate metric to evaluate the performance of NMT systems.

This paper presents the preliminary results of an ongoing academia-industry collaboration that aims to integrate MT into the workflow of Swiss Post’s Language Service. We describe the evaluations carried out to select an MT tool (commercial or open-source) and assess the suitability of machine translation for post-editing in Swiss Post’s various subject areas and language pairs. The goal of this first phase is to provide recommendations with regard to the tool, language pair and most suitable domain for implementing MT.

The project ‘Proposal and Implementation of a Swiss Research Centre for Barrier-free Communication’ (BFC) is a four-year project (2017–2020) funded by the Rectors' Conference of Swiss Higher Education Institutions (swissuniversities).1 Its purpose is to ensure that individuals with a visual or hearing disability, people with a temporary cognitive impairment and speakers without sufficient knowledge of local languages can communicate and enjoy barrier-free access to information in all spheres of life, with a special focus on higher education.

We describe CALL-SLT, a speech-enabled Computer-Assisted Language Learning application where the central idea is to prompt the student with an abstract representation of what they are supposed to say, and then use a combination of grammar-based speech recognition and rule-based translation to rate their response. The system has been developed to the level of a mature prototype, freely deployed on the web, with versions for several languages. We present an overview of the core system architecture and the various types of content we have developed. Finally, we describe several evaluations, the last of which is a study carried out over about a week using 130 subjects recruited through the Amazon Mechanical Turk, in which CALL-SLT was contrasted against a control version where the speech recognition component was disabled. The improvement in student learning performance between the two groups was significant at p < 0.02.

At present, inappropriate text alternatives for images in the Web continue to pose web accessibility barriers for people with special needs. Although research efforts have been devoted to define how to write text equivalents for visual content in websites, existing guidelines often lack direct linguistic-oriented recommendations. Similarly, most web accessibility evaluation tools just provide users with an automated functionality to check the presence of text alternatives within the element, rather than a platform to verify their content. This paper presents an overview of the findings from an exploratory study carried out to investigate if the appropriateness level of text alternatives for images in French can be improved when applying controlled language (CL) rules. Results gathered suggest that using accessibility-oriented alt style rules can have a significant impact on text alternatives’ appropriateness. Although more data would be needed to draw further conclusions about our proposal, this preliminary study already offers an interest insight into the potential use of CL checkers such as Acrolinx for language-based web accessibility evaluation.

The user-generated content represents an increasing share of the information available today. To make this type of content instantly accessible in another language, the ACCEPT project focuses on developing pre-editing technologies for correcting the source text in order to increase its translatability. Linguistically-informed pre-editing rules have been developed for English and French for the two domains considered by the project, namely, the technical domain and the healthcare domain. In this paper, we present the evaluation experiments carried out to assess the impact of the proposed pre-editing rules on translation quality. Results from a large-scale evaluation campaign show that pre-editing helps indeed attain a better translation quality for a high proportion of the data, the difference with the number of cases where the adverse effect is observed being statistically significant. The ACCEPT pre-editing technology is freely available online and can be used in any Web-based environment to enhance the translatability of user-generated content so that it reaches a broader audience.

Mismatches between training and test data are a ubiquitous problem for real SMT applications. In this paper, we examine a type of mismatch that commonly arises when translating from French and similar languages: available training data is mostly formal register, but test data may well be informal register. We consider methods for defining surface transformations that map common informal language constructions into their formal language counterparts, or vice versa; we then describe two ways to use these mappings, either to create artificial training data or to pre-process source text at run-time. An initial evaluation performed using crowd-sourced comparisons of alternate translations produced by a French-to-English SMT system suggests that both methods can improve performance, with run-time pre-processing being the more effective of the two.

We describe an experiment carried out using a French version of CALL-SLT, a web-enabled CALL game in which students at each turn are prompted to give a semi-free spoken response which the system then either accepts or rejects. The central question we investigate is whether the response is appropriate; we do this by extracting pairs of utterances where both members of the pair are responses by the same student to the same prompt, and where one response is accepted and one rejected. When the two spoken responses are presented in random order, native speakers show a reasonable degree of agreement in judging that the accepted utterance is better than the rejected one. We discuss the significance of the results and also present a small study supporting the claim that native speakers are nearly always recognised by the system, while non-native speakers are rejected a significant proportion of the time.

The goal of this paper is to provide an annotation scheme for compounds based on generative lexicon theory (GL, Pustejovsky, 1995; Bassac and Bouillon, 2001). This scheme has been tested on a set of compounds automatically extracted from the Europarl corpus (Koehn, 2005) both in Italian and French. The motivation is twofold. On the one hand, it should help refine existing compound classifications and better explain lexicalization in both languages. On the other hand, we hope that the extracted generalizations can be used in NLP, for example for improving MT systems or for query reformulation (Claveau, 2003). In this paper, we focus on the annotation scheme and its on going evaluation.

During the construction of a spoken dialogue system much effort is spent on improving the quality of speech recognition as possible. However, even if an application perfectly recognizes the input, its understanding may be far from what the user originally meant. The user should be informed about what the system actually understood so that an error will not have a negative impact in the later stages of the dialogue. One important aspect that this work tries to address is the effect of presenting the system’s understanding during interaction with users. We argue that for specific kinds of applications it’s important to confirm the understanding of the system before obtaining the output. In this way the user can avoid misconceptions and problems occurring in the dialogue flow and he can enhance his confidence in the system. Nevertheless this has an impact on the interaction, as the mental workload increases, and the user’s behavior may adapt to the system’s coverage. We focus on two applications that implement the notion of rephrasing user’s input in a different way. Our study took place among 14 subjects that used both systems on a Nokia N810 Internet Tablet.

We describe a multilingual Open Source CALL game, CALL-SLT, which reuses speech translation technology developed using the Regulus platform to create an automatic conversation partner that allows intermediate-level language students to improve their fluency. We contrast CALL-SLT with Wang's and Seneff's ``translation game'' system, in particular focussing on three issues. First, we argue that the grammar-based recognition architecture offered by Regulus is more suitable for this type of application; second, that it is preferable to prompt the student in a language-neutral form, rather than in the L1; and third, that we can profitably record successful interactions by native speakers and store them to be reused as online help for students. The current system, which will be demoed at the conference, supports four L2s (English, French, Japanese and Swedish) and two L1s (English and French). We conclude by describing an evaluation exercise, where a version of CALL-SLT configured for English L2 and French L1 was used by several hundred high school students. About half of the subjects reported positive impressions of the system.

We describe recent work on MedSLT, a medium-vocabulary interlingua-based medical speech translation system, focussing on issues that arise when handling languages of which the grammar engineer has little or no knowledge. We show how we can systematically create and maintain multiple forms of grammars, lexica and interlingual representations, with some versions being used by language informants, and some by grammar engineers. In particular, we describe the advantages of structuring the interlingua definition as a simple semantic grammar, which includes a human-readable surface form. We show how this allows us to rationalise the process of evaluating translations between languages lacking common speakers, and also makes it possible to create a simple generic tool for debugging to-interlingua translation rules. Examples presented focus on the concrete case of translation between Japanese and Arabic in both directions.

Regulus is an Open Source platform that supports construction of rule-based medium-vocabulary spoken dialogue applications. It has already been used to build several substantial speech-enabled applications, including NASA’s Clarissa procedure navigator and Geneva University’s MedSLT medical speech translator. System like these would be far more useful if they were available on a hand-held device, rather than, as with the present version, on a laptop. In this paper we describe the Open Source framework we have developed, which makes it possible to run Regulus applications on generally available mobile devices, using a distributed client-server architecture that offers transparent and reliable integration with different types of ASR systems. We describe the architecture, an implemented calendar application prototype hosted on a mobile device, and an evaluation. The evaluation shows that performance on the mobile device is as good as performance on a normal desktop PC.

Dans tout dialogue, les phrases elliptiques sont très nombreuses. Dans cet article, nous évaluons leur impact sur la reconnaissance et la traduction dans le système de traduction automatique de la parole MedSLT. La résolution des ellipses y est effectuée par une méthode robuste et portable, empruntée aux systèmes de dialogue homme-machine. Cette dernière exploite une représentation sémantique plate et combine des techniques linguistiques (pour construire la représentation) et basées sur les exemples (pour apprendre sur la base d’un corpus ce qu’est une ellipse bien formée dans un sous-domaine donné et comment la résoudre).

Cet article propose une approche originale visant la construction d’un lexique sémantique de référence sur le français. Sa principale caractéristique est de pouvoir s’appuyer sur les propriétés morphologiques des lexèmes. La méthode combine en effet des résultats d’analyse morphologique (Namer, 2002;2003), à partir de ressources lexicales de grande taille (nomenclatures du TLF) et des méthodologies d’acquisition d’information lexicale déjà éprouvées (Namer 2005; Sébillot 2002). Le format de représentation choisi, dans le cadre du Lexique Génératif, se distingue par ses propriétés d’expressivité et d’économie. Cette approche permet donc d’envisager la construction d’un lexique de référence sur le français caractérisé par une forte homogénéité tout en garantissant une couverture large, tant du point de vue de la nomenclature que du point de vue des contenus sémantiques. Une première validation de la méthode fournit une projection quantitative et qualitative des résultats attendus.

We present an overview of Regulus, an Open Source platform that supports corpus-based derivation of efficient domain-specific speech recognisers from general linguistically motivated unification grammars. We list available Open Source resources, which include compilers, resource grammars for various languages, documentation and a development environment. The greater part of the paper presents a series of experiments carried out using a medium-vocabulary medical speech translation application and a corpus of 801 recorded domain utterances, designed to investigate the impact on speech understanding performance of vocabulary size, grammatical coverage, presence or absence of various linguistic features, degree of generality of thegrammar and use or otherwise of probabilistic weighting in the CFGlanguage model. In terms of task accuracy, the most significant factors were the use of probabilistic weighting, the degree of generality of the grammar and the inclusion of features which model sortal restrictions.

Aujourd’hui, l’approche la plus courante en traitement de la parole consiste à combiner un reconnaisseur statistique avec un analyseur robuste. Pour beaucoup d’applications cependant, les reconnaisseurs linguistiques basés sur les grammaires offrent de nombreux avantages. Dans cet article, nous présentons une méthodologie et un ensemble de logiciels libres (appelé Regulus) pour dériver rapidement des reconnaisseurs linguistiquement motivés à partir d’une grammaire générale partagée pour le catalan et le français.

We present an overview of MedSLT, a medium-vocabulary medical speech translation system, focussing on the representational issues that arise when translating temporal and causal concepts. Although flat key/value structures are strongly preferred as semantic representations in speech understanding systems, we argue that it is infeasible to handle the necessary range of concepts using only flat structures. By exploiting the specific nature of the task, we show that it is possible to implement a solution which only slightly extends the representational complexity of the semantic representation language, by permitting an optional single nested level representing a subordinate clause construct. We sketch our solutions to the key problems of producing minimally nested representations using phrase-spotting methods, and writing cleanly structured rule-sets that map temporal and phrasal representations into a canonical interlingual form.

In this paper, we present evidence that providing users of a speech to speech translation system for emergency diagnosis (MedSLT) with a tool that helps them to learn the coverage greatly improves their success in using the system. In MedSLT, the system uses a grammar-based recogniser that provides more predictable results to the translation component. The help module aims at addressing the lack of robustness inherent in this type of approach. It takes as input the result of a robust statistical recogniser that performs better for out-of-coverage data and produces a list of in-coverage example sentences. These examples are selected from a defined list using a heuristic that prioritises sentences maximising the number of N-grams shared with those extracted from the recognition result.

This article describes the use and development of a tool for grammar and terminology control (FLAG), for the purposes of automating the verification of terminology for a large-scale user of multilingual terminology. It describes the various advantages of the tool and shows a process for transforming a traditional terminology list into a list of inflected forms as well as patterns which can be used to find possible morpho-syntactic derivations of terms.

This presentation outlines the implementation of a machine translation system for avalanche warning bulletins in natural language, using a unification-based formalism developed at ISSCO, which will be introduced at the same occasion. Concrete examples taken from this project exemplify a modern approach to ma- chine translation: a rich representation of the semantic content of a sentence, the use of a sin- gle grammar for parsing and generating as well as generation and transfer based exclusively on the semantic representation of a sentence. Simultaneously, the limits of bidirectional trans- fer are being tested.