Open LLMs enable AI practitioners to control development costs by building on an existing foundation for downstream applications. While offering substantial promise, current models often fail to meet the needs of users needing open solutions aligned with responsible AI principles, including data compliance, transparency, and inclusivity. In this work, we present Apertus, a fully open suite of large language models (LLMs) designed to address responsibility shortcomings in today’s open model ecosystem, namely data responsibility and global representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of data memorization, we also adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. Apertus also drastically expands multilingual coverage, training on 15T tokens from over approximately 1800 languages, with about 40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivaling or surpassing open-weight counterparts.

Annotation conflict resolution is crucial towards building machine learning models with acceptable performance. Past work on annotation conflict resolution had assumed that data is collected at once, with a fixed set of annotators and fixed annotation guidelines. Moreover, previous work dealt with atomic labeling tasks. In this paper, we address annotation conflict resolution for Natural Language Understanding (NLU), a structured prediction task, in a real-world setting of commercial voice-controlled personal assistants, where (1) regular data collections are needed to support new and existing functionalities, (2) annotation guidelines evolve over time, and (3) the pool of annotators change across data collections. We devise an approach combining information-theoretic measures and a supervised neural model to resolve conflicts in data annotation. We evaluate our approach both intrinsically and extrinsically on a real-world dataset with 3.5M utterances of a commercial dialog system in German. Our approach leads to dramatic improvements over a majority baseline especially in contentious cases. On the NLU task, our approach achieves 2.75% error reduction over a no-resolution baseline.

In this paper we report an experiment of an automated metric used to analyse the grammaticality of machine translation output. The approach (Rajman, Hartley, 2001) is based on the distribution of the linguistic information within a translated text, which is supposed similar between a learning corpus and the translation. This method is quite inexpensive, since it does not need any reference translation. First we describe the experimental method and the different tests we used. Then we show the promising results we obtained on the CESTA data, and how they correlate well with human judgments.

In this paper we present a proposal for extending the standard Wizard of Oz experimental methodology to language-enabled multimodal systems. We first discuss how Wizard of Oz experiments involving multimodal systems differ from those involving voice-only systems. We then go on to discuss the Extended Wizard of Oz methodology and the Wizard of Oz testing environment and protocol that we have developed. We then describe an example of applying this methodology to Archivus, a multimodal system for multimedia meeting retrieval and browsing. We focus in particular on the tools that the wizards would need to successfully and efficiently perform their tasks in a multimodal context. We conclude with some general comments about which questions need to be addressed when developing and using the Wizard of Oz methodology for testing multimodal systems.

This article outlines the evaluation protocol and provides the main results of the French Evaluation Campaign for Machine Translation Systems, CESTA. Following the initial objectives and evaluation plans, the evaluation metrics are briefly described: along with fluency and adequacy assessed by human judges, a number of recently proposed automated metrics are used. Two evaluation campaigns were organized, the first one in the general domain, and the second one in the medical domain. Up to six systems translating from English into French, and two systems translating from Arabic into French, took part in the campaign. The numerical results illustrate the differences between classes of systems, and provide interesting indications about the reliability of the automated metrics for French as a target language, both by comparison to human judges and using correlations between metrics. The corpora that were produced, as well as the information about the reliability of metrics, constitute reusable resources for MT evaluation.

In this paper, we report on the results of a full-size evaluation campaign of various MT systems. This campaign is novel compared to the classical DARPA/NIST MT evaluation campaigns in the sense that French is the target language, and that it includes an experiment of meta-evaluation of various metrics claiming to better predict different attributes of translation quality. We first describe the campaign, its context, its protocol and the data we used. Then we summarise the results obtained by the participating systems and discuss the meta-evaluation of the metrics used.

Les grammaires stochastiques standards utilisent des modèles probabilistes de nature générative, fondés sur des probabilités de récriture conditionnées par le symbole récrit. Les expériences montrent qu’elles tendent ainsi par nature à pénaliser les dérivations les plus longues pour une mˆeme entrée, ce qui n’est pas forcément un comportement souhaitable, ni en analyse syntaxique, ni en reconnaissance de la parole. Dans cet article, nous proposons une approche probabiliste non-générative du modèle STSG (grammaire stochastique à substitution d’arbres), selon laquelle les probabilités sont conditionnées par les feuilles des arbres syntaxiques plutˆot que par leur racine, et qui par nature fait appel à un apprentissage discriminant. Plusieurs expériences sur ce modèle sont présentées.

L’objectif de cette contribution est de présenter l’intégration de la notion d’évaluation dans la méthodologie de prototypage rapide de modèles de dialogue développée et mise en oeuvre dans le cadre du projet InfoVox. L’idée centrale de cette méthodologie est de dériver un modèle de dialogue opérationnel directement à partir du modèle de la tâche à laquelle il est associé. L’intégration systématique de différents aspects de l’évaluation dans le processus de prototypage est alors utile afin d’identifier, dès la phase de conception, les qualités et défauts de l’interface. Toutes les conclusions présentées seront illustrées par des résultats concrets obtenus au cours d’expériences réalisées dans le cadre du projet InfoVox.

L’intégration de co-occurrences dans les modèles de représentation vectorielle de documents s’est avérée une source d’amélioration de la pertinence des mesures de similarités textuelles calculées dans le cadre de ces modèles (Rajman et al., 2000; Besançon, 2001). Dans cette optique, la définition des contextes pris en compte pour les co-occurrences est cruciale, par son influence sur les performances des modèles à base de co-occurrences. Dans cet article, nous proposons d’étudier deux méthodes de filtrage des co-occurrences fondées sur l’utilisation d’informations syntaxiques supplémentaires. Nous présentons également une évaluation de ces méthodes dans le cadre de la tâche de la recherche documentaire.

Polynomial Tree Substitution Grammars, a subclass of STSGs for which finding the most probable parse is no longer NP-hard but polynomial, are defined and characterized in terms of general properties on the elementary trees in the grammar. Various sufficient and easy to compute properties for a STSG to be polynomial are presented. The min-max selection principle is shown to be one such sufficient property. In addition, another, new, instance of a sufficient property, based on lexical heads, is presented. The performances of both models are evaluated on several corpora.

The main goal of the work presented in this paper is to find an inexpensive and automatable way of predicting rankings of MT systems compatible with human evaluations of these systems expressed in the form of Fluency, Adequacy or Informativeness scores. Our approach is to establish whether there is a correlation between rankings derived from such scores and the ones that can be built on the basis of automatically computable attributes of syntactic or semantic nature. We present promising results obtained on the DARPA94 MT evaluation corpus.

Le sujet du présent article est l’intégration des sens portés par les mots en contexte dans une représentation vectorielle de textes, au moyen d’un modèle probabiliste. La représentation vectorielle considérée est le modèle DSIR, qui étend le modèle vectoriel (VS) standard en tenant compte à la fois des occurrences et des co-occurrences de mots dans les documents. L’intégration des sens dans cette représentation se fait à l’aide d’un modèle de Champ de Markov avec variables cachées, en utilisant une information sémantique dérivée de relations de synonymie extraites d’un dictionnaire de synonymes.

Trouver l’arbre d’analyse le plus probable dans le cadre du modèle DOP (Data-Oriented Parsing) — une version probabiliste de grammaire à substitution d’arbres développée par R. Bod (1992) — est connu pour être un problème NP-difficile dans le cas le plus général (Sima’an, 1996a). Cependant, si l’on introduit des restrictions a priori sur le choix des arbres élémentaires, on peut obtenir des instances particulières de DOP pour lesquelles la recherche de l’arbre d’analyse le plus probable peut être effectuée en un temps polynomial (par rapport à la taille de la phrase à analyser). La présente contribution se propose d’étudier une telle instance polynomiale de DOP, fondée sur le principe de sélection miminale-maximale et d’en évaluer les performances sur deux corpus différents.