Dissertations / Theses on the topic 'Évaluation automatique de la L2'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 50 dissertations / theses for your research on the topic 'Évaluation automatique de la L2.'
Next to every source in the list of references, there is an 'Add to bibliography' button. Press on it, and we will generate automatically the bibliographic reference to the chosen work in the citation style you need: APA, MLA, Harvard, Chicago, Vancouver, etc.
You can also download the full text of the academic publication as pdf and read online its abstract whenever available in the metadata.
Browse dissertations / theses on a wide variety of disciplines and organise your bibliography correctly.
Kobylyanskaya, Sofiya. "Towards multimodal assessment of L2 level : speech and eye tracking features in a cross-cultural setting." Electronic Thesis or Diss., université Paris-Saclay, 2024. http://www.theses.fr/2024UPASG111.
Full textIn recent years, the world of education has undergone critical changes, especially with the system’s massive digitalization in 2020, as well as the advancement of generative AI technologies. LeCycl "Learning Cyclotron" (Vargo et al., 2023) project is a part of this scientific trend and its aim is to accelerate the knowledge flow. It takes into consideration 3 main processes of learning: perception, mastering and transfer. This thesis, as part of the LeCycl project, focuses on exploring second language (L2) oral reading strategies and analyzing difficulties faced by representatives of different cultures and their techniques of coping with them. For this purpose, we are relying on multimodal cues including speech and eye tracking, as well as an original protocol that introduces nudges (represented by comic books) for the cultural adaptation (Hutin et al., 2023). For this purpose, we developed a protocol involving the collection of readingaloud data from both native and non-native English speakers (French and Japanese speakers) (Kobylyan- skaya, 2022). We analyzed speakers’ performance through acoustic and linguistic measures (phoneme realization, prosody and disfluencies such as pauses, hesitations, truncations), as well as eye movement measures (El Baha et al. 2022; Kobylyanskaya et al., 2023). Then, we used machine learning methods to define the speaker’s L2 level based on the extracted measures. Finally, we evaluate the contribution of comic books images on speakers’ oral reading performance. The results highlight that the representatives of different cultures face different challenges when reading in a foreign language and employ different strategies to overcome them, which are translated both at verbal and ocular levels. Our results underline the need for culturally adapted learning tools and the challenges involved in developing them
Leman, Adrien. "Diagnostic et évaluation automatique de la qualité vocale à partir d'indicateurs hybride." Phd thesis, INSA de Lyon, 2011. http://tel.archives-ouvertes.fr/tel-00679705.
Full textBove, Clara. "Conception et évaluation d’interfaces utilisateur explicatives pour systèmes complexes en apprentissage automatique." Electronic Thesis or Diss., Sorbonne université, 2023. https://accesdistant.sorbonne-universite.fr/login?url=https://theses-intra.sorbonne-universite.fr/2023SORUS247.pdf.
Full textThis thesis focuses on human-centered eXplainable AI (XAI) and more specif- ically on the intelligibility of Machine Learning (ML) explanations for non-expert users. The technical context is as follows: on one side, either an opaque classifier or regressor provides a prediction, with an XAI post-hoc approach that generates pieces of information as explanations; on the other side, the user receives both the prediction and the explanations. Within this XAI technical context, several is- sues might lessen the quality of explanations. The ones we focus on are: the lack of contextual information in ML explanations, the unguided design of function- alities or the user’s exploration, as well as confusion that could be caused when delivering too much information. To solve these issues, we develop an experimental procedure to design XAI functional interfaces and evaluate the intelligibility of ML explanations by non-expert users. Doing so, we investigate the XAI enhancements provided by two types of local explanation components: feature importance and counterfac- tual examples. Thus, we propose generic XAI principles for contextualizing and allowing exploration on feature importance; and for guiding users in their com- parative analysis of counterfactual explanations with plural examples. We pro- pose an implementation of such principles into two distinct explanation-based user interfaces, respectively for an insurance and a financial scenarios. Finally, we use the enhanced interfaces to conduct users studies in lab settings and to measure two dimensions of intelligibility, namely objective understanding and subjective satisfaction. For local feature importance, we demonstrate that con- textualization and exploration improve the intelligibility of such explanations. Similarly for counterfactual examples, we demonstrate that the plural condition improve the intelligibility as well, and that comparative analysis appears to be a promising tool for users’ satisfaction. At a fundamental level, we consider the issue of inconsistency within ML explanations from a theoretical point of view. In the explanation process consid- ered for this thesis, the quality of an explanation relies both on the ability of the Machine Learning system to generate a coherent explanation and on the ability of the end user to make a correct interpretation of these explanations. Thus, there can be limitations: on one side, as reported in the literature, technical limitations of ML systems might produce potentially inconsistent explanations; on the other side, human inferences can be inaccurate, even if users are presented with con- sistent explanations. Investigating such inconsistencies, we propose an ontology to structure the most common ones from the literature. We advocate that such an ontology can be useful to understand current XAI limitations for avoiding explanations pitfalls
Farenc, Christelle. "Ergoval : une méthode de structuration des règles ergonomiques permettant l'évaluation automatique d'interfaces graphiques." Toulouse 1, 1997. http://www.theses.fr/1997TOU10013.
Full textThe thesis introduces a new method for structuring ergonomic rules in order to evaluate graphical user interface. This method performed in collaboration with the SRTP (post office technical research unit) aims to be used by computer experts and to be integrated in an automatic user interface evaluation tool : ERGOVAL. In order to provide information to developers in a way they can handle it to modify the interface, ergonomic rules were reformulated to concern directly graphical objects of the user interface. Knowledge involved in the evaluation was structured in this way : * a representation of the UI in terms of the interaction objects of the norm CUA was built : this is the decomposition of graphical objects * all graphical objects concerned by the same set of ergonomic rules are grouped together into classes of objects : the typology of graphic objects. . The resulting typology consists in several levels of abstraction, the graphical objects being the leaves of this typology. The links of this typology are types of links which have hierarchical properties, i. E. Each type inherits attributes from the parent type and associated rules. A mock-up of the ERGOVAL tool was made to validate knowledge structuration and to define specifications of the final tool. In order to determine the scale application, the automatic and qualitative dimensions were studied especially the automatic retrieval of interface description and the number and level of ergonomic rules integrated in the mock-up. Consequently, the quality of an automatic evaluation and an evaluation of high level ergonomic rules were determined
Reveret, Lionel. "Conception et évaluation d'un système de suivi automatique des gestes labiaux en parole." Grenoble INPG, 1999. https://tel.archives-ouvertes.fr/tel-00389380.
Full textPho, Van-Minh. "Génération automatique de questionnaires à choix multiples pédagogiques : évaluation de l'homogénéité des options." Thesis, Paris 11, 2015. http://www.theses.fr/2015PA112192/document.
Full textRecent years have seen a revival of Intelligent Tutoring Systems. In order to make these systems widely usable by teachers and learners, they have to provide means to assist teachers in their task of exercise generation. Among these exercises, multiple-choice tests are very common. However, writing Multiple-Choice Questions (MCQ) that correctly assess a learner's level is a complex task. Guidelines were developed to manually write MCQs, but an automatic evaluation of MCQ quality would be a useful tool for teachers.We are interested in automatic evaluation of distractor (wrong answer choice) quality. To do this, we studied characteristics of relevant distractors from multiple-choice test writing guidelines. This study led us to assume that homogeneity between distractors and answer is an important criterion to validate distractors. Homogeneity is both syntactic and semantic. We validated the definition of homogeneity by a MCQ corpus analysis, and we proposed methods for automatic recognition of syntactic and semantic homogeneity based on this analysis.Then, we focused our work on distractor semantic homogeneity. To automatically estimate it, we proposed a ranking model by machine learning, combining different semantic homogeneity measures. The evaluation of the model showed that our method is more efficient than existing work to estimate distractor semantic homogeneity
Adda-Decker, Martine. "Évaluation d'unités de décision pour la reconnaissance de la parole continue." Paris 11, 1988. http://www.theses.fr/1988PA112342.
Full textVanackère, Vincent. "Trust : un système de vérification automatique de protocoles cryptographiques." Aix-Marseille 1, 2004. http://www.theses.fr/2004AIX11063.
Full textSanti, Serge. "Synthèse vocale de sons du français : modélisation acoustique et évaluation perceptive." Aix-Marseille 1, 1992. http://www.theses.fr/1992AIX10049.
Full textThis research deals with the vocal synthesis of some sounds of french. A corpus of vcv nonsense words, where v is one of the cardinal vowels a, i, u and c one of the stop consonants p, t, k, b, d, g , is considered. Relevant acoustic cues are extracted from natural speech material, then modeled by means of rules. These rules determine the evolution of the command parameters of the synthesizer. The synthetic data is evaluated by means of diagnostic evaluation methods. These evaluation tests are used not only to quantify the performance of the synthesizer but also to investigate some processes involved in both natural and synthetic speech perception. Theorical and historical aspects of speech synthesis and evaluation techniques are also included
Ciguene, Richardson. "Génération automatique de sujets d'évaluation individuels en contexte universitaire." Electronic Thesis or Diss., Amiens, 2019. http://www.theses.fr/2019AMIE0046.
Full textThis PhD work focuses on the evaluation of learning and especially the automatic generation of evaluation topics in universities. We rely on a base of source questions to create topic questions through algorithms that are able to construct differentiated assessment tests. This research has made it possible to develop a metric that measures this differentiation and to propose algorithms aimed at maximizing total differentiation on test collections, while minimizing the number of necessary patterns. The average performance of the latter depends on the number of patterns available in the source database (compared to the number of items desired in the tests), and the size of the generated collections. We focused on the possible differentiation in very small collections of subjects, and proposes methodological tracks to optimize the distribution of these differentiated subjects to cohorts of students respecting the constraints of the teacher. The rest of this work will eventually take into account the level of difficulty of a test as a new constraint, relying in part on the statistical and semantic data collected after each test. The goal is to be able to maximize the differentiation by keeping the equity between the Tests of a Collection, for an optimized distribution during the Events
Mognol, Pascal. "Contribution à la génération automatique de gammes en tournage : génération dirigée par évaluation progressive." Cachan, Ecole normale supérieure, 1994. http://www.theses.fr/1994DENS0019.
Full textCauvin, Evelyne. "Elaboration de critères prosodiques pour une évaluation semi-automatique des apprenants francophones de l'anglais." Thesis, Sorbonne Paris Cité, 2017. http://www.theses.fr/2017USPCC097/document.
Full textThe aim of our study is to modelise the prosodic interlanguage of Francophone learners of English in order to provide useful criteria for a semi-automatic assessment of their prosodic level in English. Learner assessment is a field that requires to be very rigorous and fair when setting up criteria that ensure validity, reliability, feasibility and equality, whereas English prosody is highly variable. Hence, few studies have carried out research in assessing prosody because it represents a real challenge. To address this issue, a specific strategy has been devised to elaborate a methodology that would ensure assessing a reading task successfully.The approach relies upon the constant symbiosis between prosody and a speaker’s subjective response to their environment. Our methodology, also known as « profiling », first aims at selecting relevant native perceived and acoustic prosodic features that will optimize assessment criteria by using their degree of emphasis and creating speakers’ prosodic profiles. Then, using the Longdale-Charliphonia corpus, the learner's productions are analysed acoustically. The automatic classification of the learners based on acoustic or perception prosodic variables is then submitted to expert aural assessment which assesses the learner evaluation criteria.This study achieves: A modelisation of non-native English prosody based on assessment grids that rely upon features of both native and non-native speakers of English, namely, speech rate – with or without the inclusion of pauses, register, melody and rhythm,A semi-automatic evaluation of 15 representative learners based on the above modelisation – ranking and marking,A comparison of the semi-automatic results with those of experts' auditory assessment; correspondence between the two varies from 56.83% to 59.74% when categorising the learners into three prosodic proficiency groups
Couronne, Thierry. "Analyse statistique de la performance d'un jury en évaluation sensorielle." Rennes 2, 1997. http://www.theses.fr/1997REN20009.
Full textThe aim of this thesis is to develop a methodology to analyse assessors' performance in sensory evaluation. The data used are the following : J assessors note P products on d criteria (= variables) the study of the performance is focused on panel's homogeneity. The methodology suggested uses a linear model (analysis of variance) and multidimensionnal data analysis (PCA, classification). The first part describes problematics and data : a state of the art of performance is done. The second part shows some original results. The data are analysed criterion by criterion. Usually, panel homogeneity is measured from residuals (= interaction) of the two ways model : response (of assessor J on product P) = effect of assessor J + effect of product P + residual. This work shows that, in a performance target, it is better to analyse residual of one way model with only assessor effect. This matrix of residual leads to a geometric interpretation which points out a cloud of assessors considered as variables (angles and length are of interest). First, a factorial analysis of this cloud with a contiguity constraint is introduced. Secondly, this cloud is analysed by classification. The approach of the subject leads to : - a "strict" contiguity constraint (i. E. In a class, all the couples of assessors must verify the constraint), - several agregation criteria. The document is illustrated by convenient artificial examples and by true data
Dang, Quang Vinh. "Évaluation de la confiance dans la collaboration à large échelle." Thesis, Université de Lorraine, 2018. http://www.theses.fr/2018LORR0002/document.
Full textLarge-scale collaborative systems wherein a large number of users collaborate to perform a shared task attract a lot of attention from both academic and industry. Trust is an important factor for the success of a large-scale collaboration. It is difficult for end-users to manually assess the trust level of each partner in this collaboration. We study the trust assessment problem and aim to design a computational trust model for collaborative systems. We focused on three research questions. 1. What is the effect of deploying a trust model and showing trust scores of partners to users? We designed and organized a user-experiment based on trust game, a well-known money-exchange lab-control protocol, wherein we introduced user trust scores. Our comprehensive analysis on user behavior proved that: (i) showing trust score to users encourages collaboration between them significantly at a similar level with showing nick- name, and (ii) users follow the trust score in decision-making. The results suggest that a trust model can be deployed in collaborative systems to assist users. 2. How to calculate trust score between users that experienced a collaboration? We designed a trust model for repeated trust game that computes user trust scores based on their past behavior. We validated our trust model against: (i) simulated data, (ii) human opinion, and (iii) real-world experimental data. We extended our trust model to Wikipedia based on user contributions to the quality of the edited Wikipedia articles. We proposed three machine learning approaches to assess the quality of Wikipedia articles: the first one based on random forest with manually-designed features while the other two ones based on deep learning methods. 3. How to predict trust relation between users that did not interact in the past? Given a network in which the links represent the trust/distrust relations between users, we aim to predict future relations. We proposed an algorithm that takes into account the established time information of the links in the network to predict future user trust/distrust relationships. Our algorithm outperforms state-of-the-art approaches on real-world signed directed social network datasets
Pomorski, Denis. "Apprentissage automatique symbolique/numérique : construction et évaluation d'un ensemble de règles à partir des données." Lille 1, 1991. http://www.theses.fr/1991LIL10117.
Full textBen, Jannet Mohamed Amer. "Évaluation adaptative des systèmes de transcription en contexte applicatif." Thesis, Université Paris-Saclay (ComUE), 2015. http://www.theses.fr/2015SACLS041/document.
Full textIt is important to regularly assess the technological innovation products in order to estimate the level of maturity reached by the technology and study the applications frameworks in which they can be used. Natural language processing (NLP) aims at developing modules and applications that automatically process the human language. That makes the field relevant to beth research and technological innovation. For years, the different technological modules from the NLP were developed separately. Therefore, the existing evaluation methods are in most modular. They allow to evaluate only one module at a time, while today, many applications need to combine several NLP modules to solve complex tasks. The new challenge in terms of evaluation is then to evaluate the different modules while taking into account the applicative context.Our work addresses the evaluation of Automatic Speech Recognition (ASR) systems according to the applicative context. We will focus on the case of Named Entities Recognition (NER) from spoken documents transcriped automatically. In the first part, we address the issue of evaluating ASR systems according to the application context through a study of the state of the art. We describes the tasks of ASR and NER proposed during several evalution campaigns and we discuss the protocols established for their evaluation. We also point the limitations of modular evaluation approaches and we expose the alternatives measures proposed in the literature. In the second part we describe the studied task of named entities detection, classification and decomposition and we propose a new metric ETER (Entity Tree Error Rate) which allows to take into account the specificity of the task and the applicative context during the evaluation. ETER also eliminates the biases observed with the existing metrics. In the third part, we define a new measure ATENE (Automatic Transcriptions Evaluation for Named Entities) that evaluates the quality of ASR systems and the impact of their errors for REN systems applied downstream. Rather than directly comparing reference and hypothesis transcriptions, ATENE measure how harder it becames to identify entities given the differences between hypothesis and reference by comparing an estimated likelihood of presence of entities. It is composed of two elementary measurements. The first aims to assess the risk of entities deletions and substitutions and the second aims to assess the risk of entities insertions caused by ASR errors.Our validation experiments show that the measurements given by ATENE correlate better than other measures from the state of the art with the performance of REN systems
Laaridh, Imed. "Évaluation de la parole dysarthrique : Apport du traitement automatique de la parole face à l’expertise humaine." Thesis, Avignon, 2017. http://www.theses.fr/2017AVIG0218/document.
Full textDysarthria is a speech disorder resulting from neurological impairments of the speechmotor control. It can be caused by different pathologies (Parkinson’s disease, AmyotrophicLateral Sclerosis - ALS, etc.) and affects different levels of speech production (respiratory,laryngeal and supra-laryngeal). The majority of research work dedicated tothe study of dysarthric speech relies on perceptual analyses. The most known study, byF. L. Darley in 1969, led to the organization and the classification of dysarthria within 6classes (completed with 2 additional classes in 2005).Nowadays, perceptual evaluation is still the most used method in clinical practicefor the diagnosis and the therapeutic monitoring of patients. However, this method isknown to be subjective, non reproductive and time-consuming. These limitations makeit inadequate for the evaluation of large corpora (in case of phonetic studies) or forthe follow-up of the progression of the condition of dysarthric patients. In order toovercome these limitations, professionals have been expressing their need of objectivemethods for the evaluation of disordered speech and automatic speech processing hasbeen early seen as a potential solution.The work presented in this document falls within this framework and studies thecontributions that these tools can have in the evaluation of dysarthric, and more generallypathological speech.In this work, an automatic approach for the detection of abnormal phones in dysarthricspeech is proposed and its behavior is analyzed on different speech corpora containingdifferent pathologies, dysarthric classes, dysarthria severity levels and speechstyles (read and spontaneous speech). Unlike the majority of the automatic methodsproposed in the literature that provide a global evaluation of the speech on generalitems such as dysarthria severity, intelligibility, etc., our proposed method focuses onthe phone level aiming to achieve a better characterization of dysarthria effects and toprovide a precise and useful feedback to the potential users (clinicians, phoneticians,patients). This method consists on two essential phases : (1) an automatic phone alignmentof the speech (2) an automatic classification of the resulting phones in two classes :normal and abnormal phones.When compared to an annotation of phone anomalies provided by a human expertconsidered to be the ”gold standard“, the approach showed encouraging results andproved to be able to detect anomalies on the phone level. The approach was also able to capture the evolution of the severity of the dysarthria suggesting a potential relevanceand use in the longitudinal follow-up of dysarthric patients or for the automatic predictionof their intelligibility or the severity of their dysarthria.Also, the automatic phone alignment precision was found to be dependent on the severity,the pathology, the class of the dysarthria and the phonetic category of each phone.Furthermore, the speech style was found to have an interesting effect on the behaviorsof both automatic phone alignment and anomaly detection.Finally, the results of an evaluation campaign conducted by a jury of experts on theannotations provided by the proposed approach are presented and discussed in orderto draw a panel of the strengths and limitations of the system
Pierrejean, Bénédicte. "Qualitative evaluation of word embeddings : investigating the instability in neural-based models." Thesis, Toulouse 2, 2020. http://www.theses.fr/2020TOU20001.
Full textDistributional semantics has been revolutionized by neural-based word embeddings methods such as word2vec that made semantics models more accessible by providing fast, efficient and easy to use training methods. These dense representations of lexical units based on the unsupervised analysis of large corpora are more and more used in various types of applications. They are integrated as the input layer in deep learning models or they are used to draw qualitative conclusions in corpus linguistics. However, despite their popularity, there still exists no satisfying evaluation method for word embeddings that provides a global yet precise vision of the differences between models. In this PhD thesis, we propose a methodology to qualitatively evaluate word embeddings and provide a comprehensive study of models trained using word2vec. In the first part of this thesis, we give an overview of distributional semantics evolution and review the different methods that are currently used to evaluate word embeddings. We then identify the limits of the existing methods and propose to evaluate word embeddings using a different approach based on the variation of nearest neighbors. We experiment with the proposed method by evaluating models trained with different parameters or on different corpora. Because of the non-deterministic nature of neural-based methods, we acknowledge the limits of this approach and consider the problem of nearest neighbors instability in word embeddings models. Rather than avoiding this problem we embrace it and use it as a mean to better understand word embeddings. We show that the instability problem does not impact all words in the same way and that several linguistic features are correlated. This is a step towards a better understanding of vector-based semantic models
Balmas, Françoise. "Contribution à la conceptualisation de programmes : modèle, implémentation, utilisation et évaluation." Paris 8, 1995. http://www.theses.fr/1995PA081071.
Full textTortel, Anne. "ÉVALUATION QUALITATIVE DE LA PROSODIE D'APPRENANTS FRANÇAIS: APPORT DE PARAMÉTRISATION PROSODIQUES." Phd thesis, Université de Provence - Aix-Marseille I, 2009. http://tel.archives-ouvertes.fr/tel-00455248.
Full textMdhaffar, Salima. "Reconnaissance de la parole dans un contexte de cours magistraux : évaluation, avancées et enrichissement." Thesis, Le Mans, 2020. http://www.theses.fr/2020LEMA1008.
Full textThis thesis is part of a study that explores automatic transcription potential for the instrumentation of educational situations.Our contribution covers several axes.First, we describe the enrichment and the annotation of COCo dataset that we produced as part of the ANR PASTEL project.This corpus is composed of different lectures' videos. Each lecture is related to a particular field (natural language, graphs, functions ...).In this multi-thematic framework, we are interested in the problem of the linguistic adaptation of automatic speech recognition systems (ASR).The proposed language model adaptation is based both on the lecture presentation supports provided by the teacher and in-domain data collected automatically from the web.Then, we focused on the ASR evaluation problem.The existing metrics don't allow a precise evaluation of the transcriptions' quality.Thus, we proposed two evaluation protocols.The first one deals with an intrinsic evaluation, making it possible to estimate performance only for domain words of each lecture (IWER_Average).The second protocol offers an extrinsic evaluation, which estimates the performance for two tasks exploiting transcription: information retrieval and indexability.Our experimental results show that the global word error rate (WER) masks the gain provided by language model adaptation.So, to better evaluate this gain, it seems particularly relevant to use specific measures, like those presented in this thesis.As LM adaptation is based on a collection of data from the web, we study the reproducibility of language model adaptation results by comparing the performances obtained over a long period of time.Over a collection period of one year, we were able to show that, although the data on the Web changed in part from one month to the next, the performance of the adapted transcription systems remainedconstant (i.e. no significant performance changes), no matter the period considered.Finally, we are intersted on thematic segmentation of ASR output and alignment of slides with oral lectures.For thematic segmentation, the integration of slide's change information into the TextTiling algorithm provides a significant gain in terms of F-measure.For alignment of slides with oral lectures, we have calculated a cosine similarity between the TF-IDF representation of the transcription segments andthe TF-IDF representation of text slides and we have imposed a constraint torespect the sequential order of the slides and transcription segments.Also, we have considered a confidence measure todiscuss the reliability of the proposed approach
El, Ayari Sarra. "Évaluation transparente du traitement des éléments de réponse à une question factuelle." Phd thesis, Université Paris Sud - Paris XI, 2009. http://tel.archives-ouvertes.fr/tel-00618355.
Full textRaybaud, Sylvain. "De l'utilisation de mesures de confiance en traduction automatique : évaluation, post-édition et application à la traduction de la parole." Electronic Thesis or Diss., Université de Lorraine, 2012. http://www.theses.fr/2012LORR0260.
Full textIn this thesis I shall deal with the issues of confidence estimation for machine translation and statistical machine translation of large vocabulary spontaneous speech translation. I shall first formalize the problem of confidence estimation. I present experiments under the paradigm of multivariate classification and regression. I review the performances yielded by different techniques, present the results obtained during the WMT2012 internation evaluation campaign and give the details of an application to post edition of automatically translated documents. I then deal with the issue of speech translation. After going into the details of what makes it a very specific and particularly challenging problem, I present original methods to partially solve it, by using phonetic confusion networks, confidence estimation techniques and speech segmentation. I show that the prototype I developped yields performances comparable to state-of-the-art of more standard design
Fort, Karën. "Les ressources annotées, un enjeu pour l’analyse de contenu : vers une méthodologie de l’annotation manuelle de corpus." Paris 13, 2012. http://scbd-sto.univ-paris13.fr/intranet/edgalilee_th_2012_fort.pdf.
Full textManual corpus annotation has become a key issue for Natural Langage Processing (NLP), as manually annotated corpora are used both to create and to evaluate NLP tools. However, the process of manual annotation remains underdescribed and the tools used to support it are often misused. This situation prevents the campaign manager from evaluating and guarantying the quality of the annotation. We propose in this work a unified vision of manual corpus annotation for NLP. It results from our experience of annotation campaigns, either as a manager or as a participant, as well as from collaborations with other researchers. We first propose a global methodology for managing manual corpus annotation campaigns, that relies on two pillars: an organization for annotation campaigns that puts evaluation at the heart of the process and an innovative grid for the analysis of the complexity dimensions of an annotation campaign. A second part of our work concerns the tools of the campaign manager. We evaluated the precise influence of automatic pre-annotation on the quality and speed of the correction by humans, through a series of experiments on part-of-speech tagging for English. Furthermore, we propose practical solutions for the evaluation of manual annotations, that proche che vide the campaign manager with the means to select the most appropriate measures. Finally, we brought to light the processes and tools involved in an annotation campaign and we instantiated the methodology that we described
Louis, Rambeaux Florence. "Génération et évaluation de résidus pour le diagnostic de systèmes incertains : approche fréquentielle." Nancy 1, 2001. http://www.theses.fr/2001NAN10018.
Full textDjedou, Bachir. "Prothèse cochléaire : évaluation de la perceptivité acoustique et de la séparation vocalique à travers le système chorimac." Lyon 1, 1990. http://www.theses.fr/1990LYO10114.
Full textRémillard, Judith. "Utilité et utilisation de la traduction automatique dans l’environnement de traduction : une évaluation axée sur les traducteurs professionnels." Thesis, Université d'Ottawa / University of Ottawa, 2018. http://hdl.handle.net/10393/37784.
Full textHarbaoui, Ahmed. "Vers une modélisation et un dimensionnement automatique des systèmes répartis." Phd thesis, Université de Grenoble, 2011. http://tel.archives-ouvertes.fr/tel-00649967.
Full textDang, Quang Vinh. "Évaluation de la confiance dans la collaboration à large échelle." Electronic Thesis or Diss., Université de Lorraine, 2018. http://www.theses.fr/2018LORR0002.
Full textLarge-scale collaborative systems wherein a large number of users collaborate to perform a shared task attract a lot of attention from both academic and industry. Trust is an important factor for the success of a large-scale collaboration. It is difficult for end-users to manually assess the trust level of each partner in this collaboration. We study the trust assessment problem and aim to design a computational trust model for collaborative systems. We focused on three research questions. 1. What is the effect of deploying a trust model and showing trust scores of partners to users? We designed and organized a user-experiment based on trust game, a well-known money-exchange lab-control protocol, wherein we introduced user trust scores. Our comprehensive analysis on user behavior proved that: (i) showing trust score to users encourages collaboration between them significantly at a similar level with showing nick- name, and (ii) users follow the trust score in decision-making. The results suggest that a trust model can be deployed in collaborative systems to assist users. 2. How to calculate trust score between users that experienced a collaboration? We designed a trust model for repeated trust game that computes user trust scores based on their past behavior. We validated our trust model against: (i) simulated data, (ii) human opinion, and (iii) real-world experimental data. We extended our trust model to Wikipedia based on user contributions to the quality of the edited Wikipedia articles. We proposed three machine learning approaches to assess the quality of Wikipedia articles: the first one based on random forest with manually-designed features while the other two ones based on deep learning methods. 3. How to predict trust relation between users that did not interact in the past? Given a network in which the links represent the trust/distrust relations between users, we aim to predict future relations. We proposed an algorithm that takes into account the established time information of the links in the network to predict future user trust/distrust relationships. Our algorithm outperforms state-of-the-art approaches on real-world signed directed social network datasets
Mauclair, Julie. "Mesures de confiance en traitement automatique de la parole et applications." Le Mans, 2006. http://cyberdoc.univ-lemans.fr/theses/2006/2006LEMA1027.pdf.
Full textFrérot, Cécile. "Construction et évaluation en corpus variés de lexiques syntaxiques pour la résolution des ambiguïtés de rattachement prépositionnel." Toulouse 2, 2005. http://www.theses.fr/2005TOU20048.
Full textLexicon is widely acknowledged as a very important component of any Natural Language Processing system, and the use of lexical resources is growing rapidly. Resolving Prepositional Phrase Attachment Ambiguity is known as a bottleneck in automatic parsing, and nowadays most work use corpus-based lexical resources while using existing intuition-based dictionaries is not so common. Furthermore, there has been very little work on investigating both sides (corpus and intuition-based) and measuring how each type of lexical resource helps in disambiguating. Assessing how well a lexical resource resolves Prepositional Phrase Attachment Ambiguity is mainly performed on a single corpus; therefore, very little work has been done on adapting lexical resources to the type of corpus. In our study, we build two types of corpus : one is based on an existing dictionary (Lexicon-Grammar), the other is corpus-based (a 200 million word newspaper corpus). We show how each lexicon helps in resolving Prepositional Phrase Attachment Ambiguity in five different corpora dealing with vulcanology, law, medicine, literature and journalism. We put forward some linguistic characteristics for each of the five corpora which help to understand why the performance of each lexicon varies according to the corpus. Adapting the type of lexicon resource to be used on a given corpus is made more obvious as we assess how the corpus-based lexicon performs compared with a specialised lexicon acquired from each of the five test corpora
Morlane-Hondère, François. "Une approche linguistique de l'évaluation des ressources extraites par analyse distributionnelle automatique." Phd thesis, Université Toulouse le Mirail - Toulouse II, 2013. http://tel.archives-ouvertes.fr/tel-00937926.
Full textDenoual, Etienne. "Méthodes en caractères pour le traitement automatique des langues." Phd thesis, Université Joseph Fourier (Grenoble), 2006. http://tel.archives-ouvertes.fr/tel-00107056.
Full textLe présent travail promeut l'utilisation de méthodes travaillant au niveau du signal de l'écrit: le caractère, unité immédiatement accessible dans toute langue informatisée, permet de se passer de segmentation en mots, étape actuellement incontournable pour des langues comme le chinois ou le japonais.
Dans un premier temps, nous transposons et appliquons en caractères une méthode bien établie d'évaluation objective de la traduction automatique, BLEU.
Les résultats encourageants nous permettent dans un deuxième temps d'aborder d'autres tâches de traitement des données linguistiques. Tout d'abord, le filtrage de la grammaticalité; ensuite, la caractérisation de la similarité et de l'homogénéité des ressources linguistiques. Dans toutes ces tâches, le traitement en caractères obtient des résultats acceptables, et comparables à ceux obtenus en mots.
Dans un troisième temps, nous abordons des tâches de production de données linguistiques: le calcul analogique sur les chaines de caractères permet la production de paraphrases aussi bien que la traduction automatique.
Ce travail montre qu'on peut construire un système complet de traduction automatique ne nécessitant pas de segmentation, a fortiori pour traiter des langues sans séparateur orthographique.
Bawden, Rachel. "Going beyond the sentence : Contextual Machine Translation of Dialogue." Thesis, Université Paris-Saclay (ComUE), 2018. http://www.theses.fr/2018SACLS524/document.
Full textWhile huge progress has been made in machine translation (MT) in recent years, the majority of MT systems still rely on the assumption that sentences can be translated in isolation. The result is that these MT models only have access to context within the current sentence; context from other sentences in the same text and information relevant to the scenario in which they are produced remain out of reach. The aim of contextual MT is to overcome this limitation by providing ways of integrating extra-sentential context into the translation process. Context, concerning the other sentences in the text (linguistic context) and the scenario in which the text is produced (extra-linguistic context), is important for a variety of cases, such as discourse-level and other referential phenomena. Successfully taking context into account in translation is challenging. Evaluating such strategies on their capacity to exploit context is also a challenge, standard evaluation metrics being inadequate and even misleading when it comes to assessing such improvement in contextual MT. In this thesis, we propose a range of strategies to integrate both extra-linguistic and linguistic context into the translation process. We accompany our experiments with specifically designed evaluation methods, including new test sets and corpora. Our contextual strategies include pre-processing strategies designed to disambiguate the data on which MT models are trained, post-processing strategies to integrate context by post-editing MT outputs and strategies in which context is exploited during translation proper. We cover a range of different context-dependent phenomena, including anaphoric pronoun translation, lexical disambiguation, lexical cohesion and adaptation to properties of the scenario such as speaker gender and age. Our experiments for both phrase-based statistical MT and neural MT are applied in particular to the translation of English to French and focus specifically on the translation of informal written dialogues
Raybaud, Sylvain. "De l'utilisation de mesures de confiance en traduction automatique : évaluation, post-édition et application à la traduction de la parole." Thesis, Université de Lorraine, 2012. http://www.theses.fr/2012LORR0260/document.
Full textIn this thesis I shall deal with the issues of confidence estimation for machine translation and statistical machine translation of large vocabulary spontaneous speech translation. I shall first formalize the problem of confidence estimation. I present experiments under the paradigm of multivariate classification and regression. I review the performances yielded by different techniques, present the results obtained during the WMT2012 internation evaluation campaign and give the details of an application to post edition of automatically translated documents. I then deal with the issue of speech translation. After going into the details of what makes it a very specific and particularly challenging problem, I present original methods to partially solve it, by using phonetic confusion networks, confidence estimation techniques and speech segmentation. I show that the prototype I developped yields performances comparable to state-of-the-art of more standard design
Srivastava, Brij Mohan Lal. "Anonymisation du locuteur : représentation, évaluation et garanties formelles." Thesis, Université de Lille (2018-2021), 2021. https://pepite-depot.univ-lille.fr/LIBRE/EDMADIS/2021/2021LILUB029.pdf.
Full textLarge-scale centralized storage of speech data poses severe privacy threats to the speakers. Indeed, the emergence and widespread usage of voice interfaces starting from telephone to mobile applications, and now digital assistants have enabled easier communication between the customers and the service providers. Massive speech data collection allows its users, for instance researchers, to develop tools for human convenience, like voice passwords for banking, personalized smart speakers, etc. However, centralized storage is vulnerable to cybersecurity threats which, when combined with advanced speech technologies like voice cloning, speaker recognition, and spoofing, may endow a malicious entity with the capability to re-identify speakers and breach their privacy by gaining access to their sensitive biometric characteristics, emotional states, personality attributes, pathological conditions, etc.Individuals and the members of civil society worldwide, and especially in Europe, are getting aware of this threat. With firm backing by the GDPR, several initiatives are being launched, including the publication of white papers and guidelines, to spread mass awareness and to regulate voice data so that the citizens' privacy is protected.This thesis is a timely effort to bolster such initiatives and propose solutions to remove the biometric identity of speakers from speech signals, thereby rendering them useless for re-identifying the speakers who spoke them.Besides the goal of protecting the speaker's identity from malicious access, this thesis aims to explore the solutions which do so without degrading the usefulness of speech.We present several anonymization schemes based on voice conversion methods to achieve this two-fold objective. The output of such schemes is a high-quality speech signal that is usable for publication and a variety of downstream tasks.All the schemes are subjected to a rigorous evaluation protocol which is one of the major contributions of this thesis.This protocol led to the finding that the previous approaches do not effectively protect the privacy and thereby directly inspired the VoicePrivacy initiative which is an effort to gather individuals, industry, and the scientific community to participate in building a robust anonymization scheme.We introduce a range of anonymization schemes under the purview of the VoicePrivacy initiative and empirically prove their superiority in terms of privacy protection and utility.Finally, we endeavor to remove the residual speaker identity from the anonymized speech signal using the techniques inspired by differential privacy. Such techniques provide provable analytical guarantees to the proposed anonymization schemes and open up promising perspectives for future research.In practice, the tools developed in this thesis are an essential component to build trust in any software ecosystem where voice data is stored, transmitted, processed, or published. They aim to help the organizations to comply with the rules mandated by civil governments and give a choice to individuals who wish to exercise their right to privacy
Maurel, Fabrice. "Transmodalité et multimodalité écrit/oral : modélisation, traitement automatique et évaluation de stratégies de présentation des structures "visuo-architecturale" des textes." Toulouse 3, 2004. http://www.theses.fr/2004TOU30256.
Full textWe are interested in the utility and, if the need arises, the usability of texts visual structure, within the framework of their oral transposition. We propose the synoptic of an oralisation system who leads to a text representation directly interpretable by Text-To-Speech systems. We partially realized the module specific to the oralisation strategies, in order to render some signifying parts of the text often “forgotten” by synthesis systems. The first results of this study led to specifications in the course of integration by an industrial partner. Predictive hypothesis, related to the impact on memorizing/understanding of two strategies coming from our Reformulation-based Oralisation Model for Texts Written to be Silently Read (MORTELS), have been formulated and tested. This work shows that cognitive functions was lost. Prototypes, exploiting the “Page Reflection” notion, have been conceived through interfaces in which multimodality is used to fill this gaps
Veneau, Emmanuel. "Macro-segmentation multi-critère et classification de séquences par le contenu dynamique pour l'indexation vidéo." Rennes 1, 2002. http://www.theses.fr/2002REN10013.
Full textAhmed, Assowe Houssein. "Construction et évaluation pour la TA d'un corpus journalistique bilingue : application au français-somali." Thesis, Université Grenoble Alpes (ComUE), 2019. http://www.theses.fr/2019GREAM019/document.
Full textAs part of ongoing work to computerize a large number of "poorly endowed" languages, especially those in the French-speaking world, we have created a French-Somali machine translation system dedicated to a journalistic sub-language, allowing to obtain quality translations from a bilingual body built by post-editing of GoogleTranslate results for the Somali and non-French speaking populations of the Horn of Africa. For this, we have created the very first quality French-Somali parallel corpus, comprising to date 98,912 words (about 400 standard pages) and 10,669 segments. The latter is an aligned corpus of very good quality, because we built in by post-editions editing pre-translations of produced by GT, which uses with a combination of the its French-English and English-Somali MT language pairs. It That corpus was also evaluated by 9 bilingual annotators who gave assigned a quality note score to each segment of the corpus and corrected our post-editing. From Using this growing body corpus as training corpusof work, we have built several successive versions of a MosesLIG-fr-so fragmented statistical Phrase-Based Automatic Machine Translation System (PBMT), which has proven to be better than GoogleTranslate on this language pair and this sub-language, in terms BLEU and of post-editing time. We also did used OpenNMT to build a first French-Somali neural automatic translationMT system and experiment it.in order to improve the results of TA without leading to prohibitive calculation times, both during training and during decoding.On the other hand, we have set up an iMAG (multilingual interactive access gateway) that allows non-French-speaking Somali surfers on the continent to access the online edition of the newspaper "La Nation de Djibouti" in Somali. The segments (sentences or titles), pre- automatically translated automatically by our any available fr-so MT system, can be post-edited and rated (out on a 1 to of 20scale) by the readers themselves, so as to improve the system by incremental learning, in the same way as the has been done before for the French-Chinese PBMT system. (PBMT) created by [Wang, 2015]
Le, Maguer Sébastien. "Évaluation expérimentale d'un système statistique de synthèse de la parole, HTS, pour la langue française." Phd thesis, Université Rennes 1, 2013. http://tel.archives-ouvertes.fr/tel-00934060.
Full textPillet, Constance-Aurore. "Transformation progressive du texte en données à l'aide de méthodes linguistiques et évaluation de cet apport de la linguistique sur l'efficacité du Text Mining." Paris 9, 2003. https://portail.bu.dauphine.fr/fileviewer/index.php?doc=2003PA090007.
Full textRichard, Michael. "Évaluation et validation de prévisions en loi." Thesis, Orléans, 2019. http://www.theses.fr/2019ORLE0501.
Full textIn this thesis, we study the evaluation and validation of predictive densities. In a first part, we are interested in the contribution of machine learning in the field of quantile and densityforecasting. We use some machine learning algorithms in quantile forecasting framework with real data, inorder to highlight the efficiency of particular method varying with nature of the data.In a second part, we expose some validation tests of predictive densities present in the literature. Asillustration, we use two of the mentionned tests on real data concerned about stock indexes log-returns.In the third part, we address the calibration constraint of probability forecasting. We propose a generic methodfor recalibration, which allows us to enforce this constraint. Thus, it permits to simplify the choice betweensome density forecasts. It remains to be known the impact on forecast quality, measured by predictivedistributions sharpness, or specific scores. We show that the impact on the Continuous Ranked ProbabilityScore (CRPS) is weak under some hypotheses and that it is positive under more restrictive ones. We use ourmethod on weather and electricity price ensemble forecasts.Keywords : Density forecasting, quantile forecasting, machine learning, validity tests, calibration, bias correction,PIT series , Pinball-Loss, CRPS
Colina, Zulimar. "DIAALE : Conception, implémentation et évaluation d’un dispositif en ligne d’aide à l’apprentissage de la lecture scientifique en anglais langue étrangère." Thesis, Université Clermont Auvergne (2017-2020), 2017. http://www.theses.fr/2017CLFAL021/document.
Full textThis thesis proposes to document the fields of the development of the writing comprehension in L2 among non-specialist publics and the integration of the information and communication technologies for the teaching in Venezuelan context. In order to do this, this research is based on the design, implementation and analysis of two devices of learning English as a foreign language, distant, integrating collaborative tools in the context of the faculty of engineering of the University of Carabobo in Venezuela. These two devices allow the collection of researching data and ecological data, which are structured in corpus and then processed. From these heterogeneous data, the collaborative processes between student-engineers in L2 reading tasks, the development of written comprehension skills and the use of mobile electronic devices during these collaborative tasks
Lovon, Melgarejo Jesus Enrique. "Évaluation et intégration des connaissances structurelles pour modèles de langue pré-entraînés." Electronic Thesis or Diss., Université de Toulouse (2023-....), 2024. http://thesesups.ups-tlse.fr/6065/.
Full textThe field of knowledge representation is a constantly evolving domain. Thanks to recent advancements in deep neural networks, particularly the Transformer architecture, the natural language processing (NLP) field has been provided with groundbreaking tools leading to improved performance across multiple NLP tasks. Pre-trained language models (PLMs), such as BERT and GPT, which are Transformer-based models trained on extensive amounts of textual data, have played a significant role in this progress. PLMs can produce contextualized representations embedding rich syntactic and semantic patterns of language. However, they do not provide structured and factual knowledge representations, essential for a better understanding of language. To alleviate these issues, researchers explored combining classical PLMs with external knowledge resources, such as knowledge bases (KBs). This approach aims to complement PLMs by providing the missing structural and factual components inherently present in KBs. As a result, this approach has given rise to a new family of knowledge-enhanced PLMs (KEPLMs). In this thesis, we focus on integrating KBs into PLMs, with a particular interest in their structure or hierarchy. We explore different research directions towards enhancing these PLMs, which include (i) exploring the limitations and methods to implicitly integrate KBs and their impact on reasoning-based tasks and (ii) defining evaluation methodologies for explicit hierarchical signals for PLMs and their transferability to other NLP tasks. In a first contribution, we propose to revisit the training methods of PLMs for reasoning-based tasks. Current methods are limited to generalizing this task to different difficulty levels, treating each level as a separate task. Instead, we suggest an incremental learning reasoning approach, where reasoning is learned gradually from simple to complex difficulty levels. This approach takes advantage of previously overlooked components that do not participate in the main reasoning chain, and we evaluate whether it improves the generalization of this task. We use an implicit methodology that transforms structured information into unstructured text with rich hierarchical content. We further conducted experiments on reasoning-related tasks such as reading comprehension and question answering to assess the pertinence of our proposal. For our second contribution, we aim to improve the performance of PLMs by incorporating explicit hierarchical signals into them. While various evaluation and integration approaches have been developed for static word embeddings, there is limited exploration of these methods for contextualized word embeddings. The current evaluation methods for PLMs inherit limitations from static embedding evaluations, such as dataset biases and superficial hierarchical signals. Therefore, we propose a new evaluation methodology for PLMs that considers multiple hierarchy signals. Our work characterizes the hierarchical representation by decomposing it into basic hierarchical distributions that we call hierarchy properties. We evaluate the hierarchical knowledge present in state-of-the-art PLMs using these properties and analyze if learning them aims to improve inner hierarchical representations of the models and their applicability to related NLP tasks
Lesnikova, Tatiana. "Liage de données RDF : évaluation d'approches interlingues." Thesis, Université Grenoble Alpes (ComUE), 2016. http://www.theses.fr/2016GREAM011/document.
Full textThe Semantic Web extends the Web by publishing structured and interlinked data using RDF.An RDF data set is a graph where resources are nodes labelled in natural languages. One of the key challenges of linked data is to be able to discover links across RDF data sets. Given two data sets, equivalent resources should be identified and linked by owl:sameAs links. This problem is particularly difficult when resources are described in different natural languages.This thesis investigates the effectiveness of linguistic resources for interlinking RDF data sets. For this purpose, we introduce a general framework in which each RDF resource is represented as a virtual document containing text information of neighboring nodes. The context of a resource are the labels of the neighboring nodes. Once virtual documents are created, they are projected in the same space in order to be compared. This can be achieved by using machine translation or multilingual lexical resources. Once documents are in the same space, similarity measures to find identical resources are applied. Similarity between elements of this space is taken for similarity between RDF resources.We performed evaluation of cross-lingual techniques within the proposed framework. We experimentally evaluate different methods for linking RDF data. In particular, two strategies are explored: applying machine translation or using references to multilingual resources. Overall, evaluation shows the effectiveness of cross-lingual string-based approaches for linking RDF resources expressed in different languages. The methods have been evaluated on resources in English, Chinese, French and German. The best performance (over 0.90 F-measure) was obtained by the machine translation approach. This shows that the similarity-based method can be successfully applied on RDF resources independently of their type (named entities or thesauri concepts). The best experimental results involving just a pair of languages demonstrated the usefulness of such techniques for interlinking RDF resources cross-lingually
Smits, Grégory. "Une approche par surclassement pour le contrôle d'un processus d'analyse linguistique." Caen, 2008. http://www.theses.fr/2008CAEN2014.
Full textNatural Language Processing (NLP) systems are continuously faced with the problem of generating concurrent hypotheses, of which some can be erroneous. In order to avoid the propagation of erroneous hypotheses, it appears to be essential to apply specific control strategies, which aim to distinguishing concurrent hypotheses based on their relevance. On most of observed indetermination cases, we have noticed that multiple heterogeneous knowledge sources have to be combined to determine the hypotheses relative relevance. According to this observation, we show that the control of the indetermination cases can be formalised as a decisional process based on multiple criteria. This decisional formalisation and our research of an adapted methodology have conducted us toward an outranking approach issued from the MultiCriteria Decision Aid (MCDA) paradigm. This approach differs from alternative methods by the importance granted to knowledge and preferences that an expert can express about a given problem. From this innovative intersection between NLP and MCDA, our work has been focalised on the development of a decisional module dedicated to multicriteria control. The integration of this module into a complete NLP system has allowed us to attest the feasibility of our approach and to perform experimentation on concrete indetermination cases
Nikoulina, Vassilina. "Modèle de traduction statistique à fragments enrichi par la syntaxe." Phd thesis, Grenoble, 2010. http://www.theses.fr/2010GRENM008.
Full textTraditional Statistical Machine Translation models are not aware of linguistic structure. Thus, target lexical choices and word order are controlled only by surface-based statistics learned from the training corpus. However, knowledge of linguistic structure can be beneficial since it provides generic information compensating data sparsity. The purpose of our work is to study the impact of syntactic information while preserving the general framework of Phrase-Based SMT. First, we study the integration of syntactic information using a reranking approach. We define features measuring the similarity between the dependency structures of source and target sentences, as well as features of linguistic coherence of the target sentences. The importance of each feature is assessed by learning their weights through a Structured Perceptron Algorithm. The evaluation of several reranking models shows that these features often improve the quality of translations produced by the basic model, in terms of manual evaluations as opposed to automatic measures. Then, we propose different models in order to increase the quality and diversity of the search graph produced by the decoder, through filtering out uninteresting hypotheses based on the source syntactic structure. This is done either by learning limits on the phrase recordering, or by decomposing the source sentence in order to simplify the translation process. The initial evaluations of these models look promising
Nikoulina, Vassilina. "Modèle de traduction statistique à fragments enrichi par la syntaxe." Phd thesis, Université de Grenoble, 2010. http://tel.archives-ouvertes.fr/tel-00996317.
Full textBenamar, Alexandra. "Évaluation et adaptation de plongements lexicaux au domaine à travers l'exploitation de connaissances syntaxiques et sémantiques." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG035.
Full textWord embeddings have established themselves as the most popular representation in NLP. To achieve good performance, they require training on large data sets mainly from the general domain and are frequently finetuned for specialty data. However, finetuning is a resource-intensive practice and its effectiveness is controversial.In this thesis, we evaluate the use of word embedding models on specialty corpora and show that proximity between the vocabularies of the training and application data plays a major role in the representation of out-of-vocabulary terms. We observe that this is mainly due to the initial tokenization of words and propose a measure to compute the impact of the tokenization of words on their representation. To solve this problem, we propose two methods for injecting linguistic knowledge into representations generated by Transformers: one at the data level and the other at the model level. Our research demonstrates that adding syntactic and semantic context can improve the application of self-supervised models to specialty domains, both for vocabulary representation and for NLP tasks.The proposed methods can be used for any language with linguistic information or external knowledge available. The code used for the experiments has been published to facilitate reproducibility and measures have been taken to limit the environmental impact by reducing the number of experiments
Elloumi, Zied. "Prédiction de performances des systèmes de Reconnaissance Automatique de la Parole." Thesis, Université Grenoble Alpes (ComUE), 2019. http://www.theses.fr/2019GREAM005/document.
Full textIn this thesis, we focus on performance prediction of automatic speech recognition (ASR) systems.This is a very useful task to measure the reliability of transcription hypotheses for a new data collection, when the reference transcription is unavailable and the ASR system used is unknown (black box).Our contribution focuses on several areas: first, we propose a heterogeneous French corpus to learn and evaluate ASR prediction systems.We then compare two prediction approaches: a state-of-the-art (SOTA) performance prediction based on engineered features and a new strategy based on learnt features using convolutional neural networks (CNNs).While the joint use of textual and signal features did not work for the SOTA system, the combination of inputs for CNNs leads to the best WER prediction performance. We also show that our CNN prediction remarkably predicts the shape of the WER distribution on a collection of speech recordings.Then, we analyze factors impacting both prediction approaches. We also assess the impact of the training size of prediction systems as well as the robustness of systems learned with the outputs of a particular ASR system and used to predict performance on a new data collection.Our experimental results show that both prediction approaches are robust and that the prediction task is more difficult on short speech turns as well as spontaneous speech style.Finally, we try to understand which information is captured by our neural model and its relation with different factors.Our experiences show that intermediate representations in the network automatically encode information on the speech style, the speaker's accent as well as the broadcast program type.To take advantage of this analysis, we propose a multi-task system that is slightly more effective on the performance prediction task