Dissertations / Theses on the topic 'Réseaux neuronaux (informatique) – Traitement automatique du langage naturel'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 48 dissertations / theses for your research on the topic 'Réseaux neuronaux (informatique) – Traitement automatique du langage naturel.'
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.
Jodouin, Jean-François. "Réseaux de neurones et traitement du langage naturel : étude des réseaux de neurones récurrents et de leurs représentations." Paris 11, 1993. http://www.theses.fr/1993PA112079.
Full textBardet, Adrien. "Architectures neuronales multilingues pour le traitement automatique des langues naturelles." Thesis, Le Mans, 2021. http://www.theses.fr/2021LEMA1002.
Full textThe translation of languages has become an essential need for communication between humans in a world where the possibilities of communication are expanding. Machine translation is a response to this evolving need. More recently, neural machine translation has come to the fore with the great performance of neural systems, opening up a new area of machine learning. Neural systems use large amounts of data to learn how to perform a task automatically. In the context of machine translation, the sometimes large amounts of data needed to learn efficient systems are not always available for all languages.The use of multilingual systems is one solution to this problem. Multilingual machine translation systems make it possible to translate several languages within the same system. They allow languages with little data to be learned alongside languages with more data, thus improving the performance of the translation system. This thesis focuses on multilingual machine translation approaches to improve performance for languages with limited data. I have worked on several multilingual translation approaches based on different transfer techniques between languages. The different approaches proposed, as well as additional analyses, have revealed the impact of the relevant criteria for transfer. They also show the importance, sometimes neglected, of the balance of languages within multilingual approaches
Kodelja, Bonan Dorian. "Prise en compte du contexte inter-phrastique pour l'extraction d'événements supervisée." Thesis, université Paris-Saclay, 2020. http://www.theses.fr/2020UPASS005.
Full textThe extraction of structured information from a document is one of the main parts of natural language processing (NLP). This extraction usually consists in three steps: named entities recognition relation extraction and event extraction. This last step is considered to be the most challenging. The notion of event covers a broad list of different phenomena which are characterized through a varying number of roles. Thereupon, Event extraction consists in detecting the occurrence of an event then determining its argument, that is, the different entities filling specific roles. These two steps are usually done one after the other. In this case, the first step revolves around detecting triggers indicating the occurrence of events.The current best approaches, based on neural networks, focus on the direct neighborhood of the target word in the sentence. Information in the rest of the document is then usually ignored. This thesis presents different approaches aiming at exploiting this document-level context.We begin by reproducing a state of the art convolutional neural network and analyze some of its parameters. We then present an experiment showing that, despite its good performances, our model only exploit a narrow context at the intra-sentential level.Subsequently, we present two methods to generate and integrate a representation of the inter-sentential context in a neural network operating on an intra-sentential context.The first contribution consists in producing a task-specific representation of the inter-sentential context through the aggregation of the predictions of a first intra-sentential model. This representation is then integrated in a second model, allowing it to use the document level distribution of event to improve its performances. We also show that this task-specific representation is better than an existing generic representation of the inter-sentential context.Our second contribution, in response to the limitations of the first one, allows for the dynamic generation of a specific context for each target word. This method yields the best performances for a single model on multiples datasets.Finally, we take a different tack on the exploitation of the inter-sentential context. We try a more direct modelisation of the dependencies between multiple event instances inside a document in order to produce a joint prediction. To do so, we use the PSL (Probabilistic Soft Logic) framework which allows to model such dependencies through logic formula
Ramachandra, Rao Sanjay Kamath. "Question Answering with Hybrid Data and Models." Thesis, université Paris-Saclay, 2020. http://www.theses.fr/2020UPASS024.
Full textQuestion Answering is a discipline which lies in between natural language processing and information retrieval domains. Emergence of deep learning approaches in several fields of research such as computer vision, natural language processing, speech recognition etc. has led to the rise of end-to-end models.In the context of GoASQ project, we investigate, compare and combine different approaches for answering questions formulated in natural language over textual data on open domain and biomedical domain data. The thesis work mainly focuses on 1) Building models for small scale and large scale datasets, and 2) Leveraging structured and semantic information into question answering models. Hybrid data in our research context is fusion of knowledge from free text, ontologies, entity information etc. applied towards free text question answering.The current state-of-the-art models for question answering use deep learning based models. In order to facilitate using them on small scale datasets on closed domain data, we propose to use domain adaptation. We model the BIOASQ biomedical question answering task dataset into two different QA task models and show how the Open Domain Question Answering task suits better than the Reading Comprehension task by comparing experimental results. We pre-train the Reading Comprehension model with different datasets to show the variability in performance when these models are adapted to biomedical domain. We find that using one particular dataset (SQUAD v2.0 dataset) for pre-training performs the best on single dataset pre-training and a combination of four Reading Comprehension datasets performed the best towards the biomedical domain adaptation. We perform some of the above experiments using large scale pre-trained language models like BERT which are fine-tuned to the question answering task. The performance varies based on the type of data used to pre-train BERT. For BERT pre-training on the language modelling task, we find the biomedical data trained BIOBERT to be the best choice for biomedical QA.Since deep learning models tend to function in an end-to-end fashion, semantic and structured information coming from expert annotated information sources are not explicitly used. We highlight the necessity for using Lexical and Expected Answer Types in open domain and biomedical domain question answering by performing several verification experiments. These types are used to highlight entities in two QA tasks which shows improvements while using entity embeddings based on the answer type annotations. We manually annotated an answer variant dataset for BIOASQ and show the importance of learning a QA model with answer variants present in the paragraphs.Our hypothesis is that the results obtained from deep learning models can further be improved using semantic features and collective features from different paragraphs for a question. We propose to use ranking models based on binary classification methods to better rank Top-1 prediction among Top-K predictions using these features, leading to an hybrid model that outperforms state-of-art-results on several datasets. We experiment with several overall Open Domain Question Answering models on QA sub-task datasets built for Reading Comprehension and Answer Sentence Selection tasks. We show the difference in performance when these are modelled as overall QA task and highlight the wide gap in building end-to-end models for overall question answering task
Janod, Killian. "La représentation des documents par réseaux de neurones pour la compréhension de documents parlés." Thesis, Avignon, 2017. http://www.theses.fr/2017AVIG0222/document.
Full textApplication of spoken language understanding aim to extract relevant items of meaning from spoken signal. There is two distinct types of spoken language understanding : understanding of human/human dialogue and understanding in human/machine dialogue. Given a type of conversation, the structure of dialogues and the goal of the understanding process varies. However, in both cases, most of the time, automatic systems have a step of speech recognition to generate the textual transcript of the spoken signal. Speech recognition systems in adverse conditions, even the most advanced one, produce erroneous or partly erroneous transcript of speech. Those errors can be explained by the presence of information of various natures and functions such as speaker and ambience specificities. They can have an important adverse impact on the performance of the understanding process. The first part of the contribution in this thesis shows that using deep autoencoders produce a more abstract latent representation of the transcript. This latent representation allow spoken language understanding system to be more robust to automatic transcription mistakes. In the other part, we propose two different approaches to generate more robust representation by combining multiple views of a given dialogue in order to improve the results of the spoken language understanding system. The first approach combine multiple thematic spaces to produce a better representation. The second one introduce new autoencoders architectures that use supervision in the denoising autoencoders. These contributions show that these architectures reduce the difference in performance between a spoken language understanding using automatic transcript and one using manual transcript
Petit, Alban. "Structured prediction methods for semantic parsing." Electronic Thesis or Diss., université Paris-Saclay, 2024. http://www.theses.fr/2024UPASG002.
Full textSemantic parsing is the task of mapping a natural language utterance into a formal representation that can be manipulated by a computer program. It is a major task in Natural Language Processing with several applications, including the development of questions answers systems or code generation among others.In recent years, neural-based approaches and particularly sequence-to-sequence architectures have demonstrated strong performances on this task. However, several works have put forward the limitations of neural-based parsers on out-of-distribution examples. In particular, they fail when compositional generalization is required. It is thus essential to develop parsers that exhibit better compositional abilities.The representation of the semantic content is another concern when tackling semantic parsing. As different syntactic structures can be used to represent the same semantic content, one should focus on structures that can both accurately represent the semantic content and align well with natural language. In that regard, this thesis relies on graph-based representations for semantic parsing and focuses on two tasks.The first one deals with the training of graph-based semantic parsers. They need to learn a correspondence between the parts of the semantic graph and the natural language utterance. As this information is usually absent in the training data, we propose training algorithms that treat this correspondence as a latent variable.The second task focuses on improving the compositional abilities of graph-based semantic parsers in two different settings. Note that in graph prediction, the traditional pipeline is to first predict the nodes and then the arcs of the graph. In the first setting, we assume that the graphs that must be predicted are trees and propose an optimization algorithm based on constraint smoothing and conditional gradient that allows to predict the entire graph jointly. In the second setting, we do not make any assumption regarding the nature of the semantic graphs. In that case, we propose to introduce an intermediate supertagging step in the inference pipeline that constrains the arc prediction step. In both settings, our contributions can be viewed as introducing additional local constraints to ensure the well-formedness the overall prediction. Experimentally, our contributions significantly improve the compositional abilities of graph-based semantic parsers and outperform comparable baselines on several datasets designed to evaluate compositional generalization
Ngo, Ho Anh Khoa. "Generative Probabilistic Alignment Models for Words and Subwords : a Systematic Exploration of the Limits and Potentials of Neural Parametrizations." Electronic Thesis or Diss., université Paris-Saclay, 2021. http://www.theses.fr/2021UPASG014.
Full textAlignment consists of establishing a mapping between units in a bitext, combining a text in a source language and its translation in a target language. Alignments can be computed at several levels: between documents, between sentences, between phrases, between words, or even between smaller units end when one of the languages is morphologically complex, which implies to align fragments of words (morphemes). Alignments can also be considered between more complex linguistic structures such as trees or graphs. This is a complex, under-specified task that humans accomplish with difficulty. Its automation is a notoriously difficult problem in natural language processing, historically associated with the first probabilistic word-based translation models. The design of new models for natural language processing, based on distributed representations computed by neural networks, allows us to question and revisit the computation of these alignments. This research project, therefore, aims to comprehensively understand the limitations of existing statistical alignment models and to design neural models that can be learned without supervision to overcome these drawbacks and to improve the state of art in terms of alignment accuracy
Parcollet, Titouan. "Quaternion neural networks A survey of quaternion neural networks - Chapter 2 Real to H-space Autoencoders for Theme Identification in Telephone Conversations - Chapter 7." Thesis, Avignon, 2019. http://www.theses.fr/2019AVIG0233.
Full textIn the recent years, deep learning has become the leading approach to modern artificial intelligence (AI). The important improvement in terms of processing time required for learning AI based models alongside with the growing amount of available data made of deep neural networks (DNN) the strongest solution to solve complex real-world problems. However, a major challenge of artificial neural architectures lies on better considering the high-dimensionality of the data.To alleviate this issue, neural networks (NN) based on complex and hypercomplex algebras have been developped. The natural multidimensionality of the data is elegantly embedded within complex and hypercomplex neurons composing the model. In particular, quaternion neural networks (QNN) have been proposed to deal with up to four dimensional features, based on the quaternion representation of rotations and orientations. Unfortunately, and conversely to complex-valued neural networks that are nowadays known as a strong alternative to real-valued neural networks, QNNs suffer from numerous limitations that are carrefuly addressed in the different parts detailled in this thesis.The thesis consists in three parts that gradually introduce the missing concepts of QNNs, to make them a strong alternative to real-valued NNs. The first part introduces and list previous findings on quaternion numbers and quaternion neural networks to define the context and strong basics for building elaborated QNNs.The second part introduces state-of-the-art quaternion neural networks for a fair comparison with real-valued neural architectures. More precisely, QNNs were limited by their simple architectures that were mostly composed of a single and shallow hidden layer. In this part, we propose to bridge the gap between quaternion and real-valued models by presenting different quaternion architectures. First, basic paradigms such as autoencoders and deep fully-connected neural networks are introduced. Then, more elaborated convolutional and recurrent neural networks are extended to the quaternion domain. Experiments to compare QNNs over equivalents NNs have been conducted on real-world tasks across various domains, including computer vision, spoken language understanding and speech recognition. QNNs increase performances while reducing the needed number of neural parameters compared to real-valued neural networks.Then, QNNs are extended to unconventional settings. In a conventional QNN scenario, input features are manually segmented into three or four components, enabling further quaternion processing. Unfortunately, there is no evidence that such manual segmentation is the representation that suits the most to solve the considered task. Morevover, a manual segmentation drastically reduces the field of application of QNNs to four dimensional use-cases. Therefore the third part introduces a supervised and an unsupervised model to extract meaningful and disantengled quaternion input features, from any real-valued input signal, enabling the use of QNNs regardless of the dimensionality of the considered task. Conducted experiments on speech recognition and document classification show that the proposed approaches outperform traditional quaternion features
Tafforeau, Jérémie. "Modèle joint pour le traitement automatique de la langue : perspectives au travers des réseaux de neurones." Thesis, Aix-Marseille, 2017. http://www.theses.fr/2017AIXM0430/document.
Full textNLP researchers has identified different levels of linguistic analysis. This lead to a hierarchical division of the various tasks performed in order to analyze a text statement. The traditional approach considers task-specific models which are subsequently arranged in cascade within processing chains (pipelines). This approach has a number of limitations: the empirical selection of models features, the errors accumulation in the pipeline and the lack of robusteness to domain changes. These limitations lead to particularly high performance losses in the case of non-canonical language with limited data available such as transcriptions of conversations over phone. Disfluencies and speech-specific syntactic schemes, as well as transcription errors in automatic speech recognition systems, lead to a significant drop of performances. It is therefore necessary to develop robust and flexible systems. We intend to perform a syntactic and semantic analysis using a deep neural network multitask model while taking into account the variations of domain and/or language registers within the data
Piat, Guilhem Xavier. "Incorporating expert knowledge in deep neural networks for domain adaptation in natural language processing." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG087.
Full textCurrent state-of-the-art Language Models (LMs) are able to converse, summarize, translate, solve novel problems, reason, and use abstract concepts at a near-human level. However, to achieve such abilities, and in particular to acquire ``common sense'' and domain-specific knowledge, they require vast amounts of text, which are not available in all languages or domains. Additionally, their computational requirements are out of reach for most organizations, limiting their potential for specificity and their applicability in the context of sensitive data.Knowledge Graphs (KGs) are sources of structured knowledge which associate linguistic concepts through semantic relations. These graphs are sources of high quality knowledge which pre-exist in a variety of otherwise low-resource domains, and are denser in information than typical text. By allowing LMs to leverage these information structures, we could remove the burden of memorizing facts from LMs, reducing the amount of text and computation required to train them and allowing us to update their knowledge with little to no additional training by updating the KGs, therefore broadening their scope of applicability and making them more democratizable.Various approaches have succeeded in improving Transformer-based LMs using KGs. However, most of them unrealistically assume the problem of Entity Linking (EL), i.e. determining which KG concepts are present in the text, is solved upstream. This thesis covers the limitations of handling EL as an upstream task. It goes on to examine the possibility of learning EL jointly with language modeling, and finds that while this is a viable strategy, it does little to decrease the LM's reliance on in-domain text. Lastly, this thesis covers the strategy of using KGs to generate text in order to leverage LMs' linguistic abilities and finds that even naïve implementations of this approach can result in measurable improvements on in-domain language processing
Delecraz, Sébastien. "Approches jointes texte/image pour la compréhension multimodale de documents." Electronic Thesis or Diss., Aix-Marseille, 2018. http://www.theses.fr/2018AIXM0634.
Full textThe human faculties of understanding are essentially multimodal. To understand the world around them, human beings fuse the information coming from all of their sensory receptors. Most of the documents used in automatic information processing contain multimodal information, for example text and image in textual documents or image and sound in video documents, however the processings used are most often monomodal. The aim of this thesis is to propose joint processes applying mainly to text and image for the processing of multimodal documents through two studies: one on multimodal fusion for the speaker role recognition in television broadcasts, the other on the complementarity of modalities for a task of linguistic analysis on corpora of images with captions. In the first part of this study, we interested in audiovisual documents analysis from news television channels. We propose an approach that uses in particular deep neural networks for representation and fusion of modalities. In the second part of this thesis, we are interested in approaches allowing to use several sources of multimodal information for a monomodal task of natural language processing in order to study their complementarity. We propose a complete system of correction of prepositional attachments using visual information, trained on a multimodal corpus of images with captions
Strub, Florian. "Développement de modèles multimodaux interactifs pour l'apprentissage du langage dans des environnements visuels." Thesis, Lille 1, 2020. http://www.theses.fr/2020LIL1I030.
Full textWhile our representation of the world is shaped by our perceptions, our languages, and our interactions, they have traditionally been distinct fields of study in machine learning. Fortunately, this partitioning started opening up with the recent advents of deep learning methods, which standardized raw feature extraction across communities. However, multimodal neural architectures are still at their beginning, and deep reinforcement learning is often limited to constrained environments. Yet, we ideally aim to develop large-scale multimodal and interactive models towards correctly apprehending the complexity of the world. As a first milestone, this thesis focuses on visually grounded language learning for three reasons (i) they are both well-studied modalities across different scientific fields (ii) it builds upon deep learning breakthroughs in natural language processing and computer vision (ii) the interplay between language and vision has been acknowledged in cognitive science. More precisely, we first designed the GuessWhat?! game for assessing visually grounded language understanding of the models: two players collaborate to locate a hidden object in an image by asking a sequence of questions. We then introduce modulation as a novel deep multimodal mechanism, and we show that it successfully fuses visual and linguistic representations by taking advantage of the hierarchical structure of neural networks. Finally, we investigate how reinforcement learning can support visually grounded language learning and cement the underlying multimodal representation. We show that such interactive learning leads to consistent language strategies but gives raise to new research issues
Francis, Danny. "Représentations sémantiques d'images et de vidéos." Electronic Thesis or Diss., Sorbonne université, 2019. http://www.theses.fr/2019SORUS605.
Full textRecent research in Deep Learning has sent the quality of results in multimedia tasks rocketing: thanks to new big datasets of annotated images and videos, Deep Neural Networks (DNN) have outperformed other models in most cases. In this thesis, we aim at developing DNN models for automatically deriving semantic representations of images and videos. In particular we focus on two main tasks : vision-text matching and image/video automatic captioning. Addressing the matching task can be done by comparing visual objects and texts in a visual space, a textual space or a multimodal space. Based on recent works on capsule networks, we define two novel models to address the vision-text matching problem: Recurrent Capsule Networks and Gated Recurrent Capsules. In image and video captioning, we have to tackle a challenging task where a visual object has to be analyzed, and translated into a textual description in natural language. For that purpose, we propose two novel curriculum learning methods. Moreover regarding video captioning, analyzing videos requires not only to parse still images, but also to draw correspondences through time. We propose a novel Learned Spatio-Temporal Adaptive Pooling method for video captioning that combines spatial and temporal analysis. Extensive experiments on standard datasets assess the interest of our models and methods with respect to existing works
Xiao, Chunyang. "Neural-Symbolic Learning for Semantic Parsing." Thesis, Université de Lorraine, 2017. http://www.theses.fr/2017LORR0268/document.
Full textOur goal in this thesis is to build a system that answers a natural language question (NL) by representing its semantics as a logical form (LF) and then computing the answer by executing the LF over a knowledge base. The core part of such a system is the semantic parser that maps questions to logical forms. Our focus is how to build high-performance semantic parsers by learning from (NL, LF) pairs. We propose to combine recurrent neural networks (RNNs) with symbolic prior knowledge expressed through context-free grammars (CFGs) and automata. By integrating CFGs over LFs into the RNN training and inference processes, we guarantee that the generated logical forms are well-formed; by integrating, through weighted automata, prior knowledge over the presence of certain entities in the LF, we further enhance the performance of our models. Experimentally, we show that our approach achieves better performance than previous semantic parsers not using neural networks as well as RNNs not informed by such prior knowledge
Veron, Mathilde. "Systèmes de dialogue apprenant tout au long de leur vie : de l'élaboration à l'évaluation." Electronic Thesis or Diss., université Paris-Saclay, 2022. http://www.theses.fr/2022UPASG089.
Full textTask-oriented dialogue systems, more commonly known as chatbots, are intended to perform tasks and provide the information required by a user in a conversation in a specific domain (e.g., train booking). These systems have been widely adopted by many companies. However, they suffer in practice from some limitations: (1) they are dependent on the training data needed to obtain a performing system, (2) they lack flexibility and perform poorly as soon as the case encountered in practice moves away from the data seen during development, and (3) it is difficult to adapt them over time to new elements that appear given the inevitable evolution of the world, of the requirements of the designers and users. Thus, we apply Lifelong Learning (LL) to task-oriented dialogue systems. We define LL as the ability of a system to be applied to and learn multiple tasks over time, in production, autonomously, continuously, and interactively. Three steps must be performed in autonomy by the system: (1) Detect the presence of a new element, (2) extract and identify the new element, and (3) adapt the system components associated with this element. As part of this thesis and given the complexity of LL, we focus our work on three subproblems associated with LL dialogue systems. As a first step, we propose a first methodology for the continuous and time-dependent evaluation of on-the-job learning dialogue systems. This type of learning is close to LL but puts aside the multitask aspect. We also describe a task-oriented dialogue system capable of improving its slot detection on-the-job via the autonomous annotation of data collected during its interactions. We evaluate this system through two adaptation methods using our methodology and show interest in a continuous evaluation over time. As a second step, we focus on the innovative study of interlingual transfer when applying continual learning to a language sequence. Indeed, transfer and continual learning are two main aspects of LL. We perform this study on the slot-filling task using multilingual BERT. We observe substantial forward transfer capabilities despite the presence of forgetting and demonstrate the capabilities of a model trained in a continual manner. As a third step, we study inter-domain transfer in the context of zero-shot learning. We carry out this study on a task that requires considering the whole dialogue and not only the current turn, which corresponds to the dialogue state tracking task. We first study the generalization and transfer capabilities of an existing model on new slot values. Then, we propose some model variants and a method able to improve the zero-shot performance of the model on new types of slots belonging to a new domain
Ortiz, Suarez Pedro. "A Data-driven Approach to Natural Language Processing for Contemporary and Historical French." Electronic Thesis or Diss., Sorbonne université, 2022. http://www.theses.fr/2022SORUS155.
Full textIn recent years, neural methods for Natural Language Processing (NLP) have consistently and repeatedly improved the state of the art in a wide variety of NLP tasks. One of the main contributing reasons for this steady improvement is the increased use of transfer learning techniques. These methods consist in taking a pre-trained model and reusing it, with little to no further training, to solve other tasks. Even though these models have clear advantages, their main drawback is the amount of data that is needed to pre-train them. The lack of availability of large-scale data previously hindered the development of such models for contemporary French, and even more so for its historical states.In this thesis, we focus on developing corpora for the pre-training of these transfer learning architectures. This approach proves to be extremely effective, as we are able to establish a new state of the art for a wide range of tasks in NLP for contemporary, medieval and early modern French as well as for six other contemporary languages. Furthermore, we are able to determine, not only that these models are extremely sensitive to pre-training data quality, heterogeneity and balance, but we also show that these three features are better predictors of the pre-trained models' performance in downstream tasks than the pre-training data size itself. In fact, we determine that the importance of the pre-training dataset size was largely overestimated, as we are able to repeatedly show that such models can be pre-trained with corpora of a modest size
Cadène, Rémi. "Deep Multimodal Learning for Vision and Language Processing." Electronic Thesis or Diss., Sorbonne université, 2020. http://www.theses.fr/2020SORUS277.
Full textDigital technologies have become instrumental in transforming our society. Recent statistical methods have been successfully deployed to automate the processing of the growing amount of images, videos, and texts we produce daily. In particular, deep neural networks have been adopted by the computer vision and natural language processing communities for their ability to perform accurate image recognition and text understanding once trained on big sets of data. Advances in both communities built the groundwork for new research problems at the intersection of vision and language. Integrating language into visual recognition could have an important impact on human life through the creation of real-world applications such as next-generation search engines or AI assistants.In the first part of this thesis, we focus on systems for cross-modal text-image retrieval. We propose a learning strategy to efficiently align both modalities while structuring the retrieval space with semantic information. In the second part, we focus on systems able to answer questions about an image. We propose a multimodal architecture that iteratively fuses the visual and textual modalities using a factorized bilinear model while modeling pairwise relationships between each region of the image. In the last part, we address the issues related to biases in the modeling. We propose a learning strategy to reduce the language biases which are commonly present in visual question answering systems
Benamar, 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
Labeau, Matthieu. "Neural language models : Dealing with large vocabularies." Thesis, Université Paris-Saclay (ComUE), 2018. http://www.theses.fr/2018SACLS313/document.
Full textThis work investigates practical methods to ease training and improve performances of neural language models with large vocabularies. The main limitation of neural language models is their expensive computational cost: it depends on the size of the vocabulary, with which it grows linearly. Despite several training tricks, the most straightforward way to limit computation time is to limit the vocabulary size, which is not a satisfactory solution for numerous tasks. Most of the existing methods used to train large-vocabulary language models revolve around avoiding the computation of the partition function, ensuring that output scores are normalized into a probability distribution. Here, we focus on sampling-based approaches, including importance sampling and noise contrastive estimation. These methods allow an approximate computation of the partition function. After examining the mechanism of self-normalization in noise-contrastive estimation, we first propose to improve its efficiency with solutions that are adapted to the inner workings of the method and experimentally show that they considerably ease training. Our second contribution is to expand on a generalization of several sampling based objectives as Bregman divergences, in order to experiment with new objectives. We use Beta divergences to derive a set of objectives from which noise contrastive estimation is a particular case. Finally, we aim at improving performances on full vocabulary language models, by augmenting output words representation with subwords. We experiment on a Czech dataset and show that using character-based representations besides word embeddings for output representations gives better results. We also show that reducing the size of the output look-up table improves results even more
Bannour, Nesrine. "Information Extraction from Electronic Health Records : Studies on temporal ordering, privacy and environmental impact." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG082.
Full textAutomatically extracting rich information contained in Electronic Health Records (EHRs) is crucial to improve clinical research. However, most of this information is in the form of unstructured text.The complexity and the sensitive nature of clinical text involve further challenges. As a result, sharing data is difficult in practice and is governed by regulations. Neural-based models showed impressive results for Information Extraction, but they need significant amounts of manually annotated data, which is often limited, particularly for non-English languages. Thus, the performance is still not ideal for practical use. In addition to privacy issues, using deep learning models has a significant environmental impact.In this thesis, we develop methods and resources for clinical Named Entity Recognition (NER) and Temporal Relation Extraction (TRE) in French clinical narratives.Specifically, we propose a privacy-preserving mimic models architecture by exploring the mimic learning approach to enable knowledge transfer through a teacher model trained on a private corpus to a student model. This student model could be publicly shared without disclosing the original sensitive data or the private teacher model on which it was trained. Our strategy offers a good compromise between performance and data privacy preservation.Then, we introduce a novel event- and task-independent representation of temporal relations. Our representation enables identifying homogeneous text portions from a temporal standpoint and classifying the relation between each text portion and the document creation time. This makes the annotation and extraction of temporal relations easier and reproducible through different event types, as no prior definition and extraction of events is required.Finally, we conduct a comparative analysis of existing tools for measuring the carbon emissions of NLP models. We adopt one of the studied tools to calculate the carbon footprint of all our created models during the thesis, as we consider it a first step toward increasing awareness and control of their environmental impact.To summarize, we generate shareable privacy-preserving NER models that clinicians can efficiently use. We also demonstrate that the TRE task may be tackled independently of the application domain and that good results can be obtained using real-world oncology clinical notes
Barhoumi, Amira. "Une approche neuronale pour l’analyse d’opinions en arabe." Thesis, Le Mans, 2020. http://www.theses.fr/2020LEMA1022.
Full textMy thesis is part of Arabic sentiment analysis. Its aim is to determine the global polarity of a given textual statement written in MSA or dialectal arabic. This research area has been subject of numerous studies dealing with Indo-European languages, in particular English. One of difficulties confronting this thesis is the processing of Arabic. In fact, Arabic is a morphologically rich language which implies a greater sparsity : we want to overcome this problem by producing, in a completely automatic way, new arabic specific embeddings. Our study focuses on the use of a neural approach to improve polarity detection, using embeddings. These embeddings have revealed fundamental in various natural languages processing tasks (NLP). Our contribution in this thesis concerns several axis. First, we begin with a preliminary study of the various existing pre-trained word embeddings resources in arabic. These embeddings consider words as space separated units in order to capture semantic and syntactic similarities in the embedding space. Second, we focus on the specifity of Arabic language. We propose arabic specific embeddings that take into account agglutination and morphological richness of Arabic. These specific embeddings have been used, alone and in combined way, as input to neural networks providing an improvement in terms of classification performance. Finally, we evaluate embeddings with intrinsic and extrinsic methods specific to sentiment analysis task. For intrinsic embeddings evaluation, we propose a new protocol introducing the notion of sentiment stability in the embeddings space. We propose also a qualitaive extrinsic analysis of our embeddings by using visualisation methods
Delecraz, Sébastien. "Approches jointes texte/image pour la compréhension multimodale de documents." Thesis, Aix-Marseille, 2018. http://www.theses.fr/2018AIXM0634/document.
Full textThe human faculties of understanding are essentially multimodal. To understand the world around them, human beings fuse the information coming from all of their sensory receptors. Most of the documents used in automatic information processing contain multimodal information, for example text and image in textual documents or image and sound in video documents, however the processings used are most often monomodal. The aim of this thesis is to propose joint processes applying mainly to text and image for the processing of multimodal documents through two studies: one on multimodal fusion for the speaker role recognition in television broadcasts, the other on the complementarity of modalities for a task of linguistic analysis on corpora of images with captions. In the first part of this study, we interested in audiovisual documents analysis from news television channels. We propose an approach that uses in particular deep neural networks for representation and fusion of modalities. In the second part of this thesis, we are interested in approaches allowing to use several sources of multimodal information for a monomodal task of natural language processing in order to study their complementarity. We propose a complete system of correction of prepositional attachments using visual information, trained on a multimodal corpus of images with captions
Zablocki, Éloi. "Multimodal machine learning : complementarity of textual and visual contexts." Electronic Thesis or Diss., Sorbonne université, 2019. http://www.theses.fr/2019SORUS409.
Full textResearch looking at the interaction between language and vision, despite a growing interest, is relatively underexplored. Beyond trivial differences between texts and images, these two modalities have non overlapping semantics. On the one hand, language can express high-level semantics about the world, but it is biased in the sense that a large portion of its content is implicit (common-sense or implicit knowledge). On the other hand, images are aggregates of lower-level information, but they can depict a more direct view of real-world statistics and can be used to ground the meaning of objects. In this thesis, we exploit connections and leverage complementarity between language and vision. First, natural language understanding capacities can be augmented with the help of the visual modality, as language is known to be grounded in the visual world. In particular, representing language semantics is a long-standing problem for the natural language processing community, and to further improve traditional approaches towards that goal, leveraging visual information is crucial. We show that semantic linguistic representations can be enriched by visual information, and we especially focus on visual contexts and spatial organization of scenes. We present two models to learn grounded word or sentence semantic representations respectively, with the help of images. Conversely, integrating language with vision brings the possibility of expanding the horizons and tasks of the vision community. Assuming that language contains visual information about objects, and that this can be captured within linguistic semantic representation, we focus on the zero-shot object recognition task, which consists in recognizing objects that have never been seen thanks to linguistic knowledge acquired about the objects beforehand. In particular, we argue that linguistic representations not only contain visual information about the visual appearance of objects but also about their typical visual surroundings and visual occurrence frequencies. We thus present a model for zero-shot recognition that leverages the visual context of an object, and its visual occurrence likelihood, in addition to the region of interest as done in traditional approaches. Finally, we present prospective research directions to further exploit connections between language and images and to better understand the semantic gap between the two modalities
Montariol, Syrielle. "Models of diachronic semantic change using word embeddings." Electronic Thesis or Diss., université Paris-Saclay, 2021. http://www.theses.fr/2021UPASG006.
Full textIn this thesis, we study lexical semantic change: temporal variations in the use and meaning of words, also called extit{diachrony}. These changes are carried by the way people use words, and mirror the evolution of various aspects of society such as its technological and cultural environment.We explore, compare and evaluate methods to build time-varying embeddings from a corpus in order to analyse language evolution.We focus on contextualised word embeddings using pre-trained language models such as BERT. We propose several approaches to extract and aggregate the contextualised representations of words over time, and quantify their level of semantic change.In particular, we address the practical aspect of these systems: the scalability of our approaches, with a view to applying them to large corpora or large vocabularies; their interpretability, by disambiguating the different uses of a word over time; and their applicability to concrete issues, for documents related to COVID19We evaluate the efficiency of these methods quantitatively using several annotated corpora, and qualitatively by linking the detected semantic variations with real-life events and numerical data.Finally, we extend the task of semantic change detection beyond the temporal dimension. We adapt it to a bilingual setting, to study the joint evolution of a word and its translation in two corpora of different languages; and to a synchronic frame, to detect semantic variations across different sources or communities on top of the temporal variation
Meftah, Sara. "Neural Transfer Learning for Domain Adaptation in Natural Language Processing." Thesis, université Paris-Saclay, 2021. http://www.theses.fr/2021UPASG021.
Full textRecent approaches based on end-to-end deep neural networks have revolutionised Natural Language Processing (NLP), achieving remarkable results in several tasks and languages. Nevertheless, these approaches are limited with their "gluttony" in terms of annotated data, since they rely on a supervised training paradigm, i.e. training from scratch on large amounts of annotated data. Therefore, there is a wide gap between NLP technologies capabilities for high-resource languages compared to the long tail of low-resourced languages. Moreover, NLP researchers have focused much of their effort on training NLP models on the news domain, due to the availability of training data. However, many research works have highlighted that models trained on news fail to work efficiently on out-of-domain data, due to their lack of robustness against domain shifts. This thesis presents a study of transfer learning approaches, through which we propose different methods to take benefit from the pre-learned knowledge on the high-resourced domain to enhance the performance of neural NLP models in low-resourced settings. Precisely, we apply our approaches to transfer from the news domain to the social media domain. Indeed, despite the importance of its valuable content for a variety of applications (e.g. public security, health monitoring, or trends highlight), this domain is still poor in terms of annotated data. We present different contributions. First, we propose two methods to transfer the knowledge encoded in the neural representations of a source model pretrained on large labelled datasets from the source domain to the target model, further adapted by a fine-tuning on few annotated examples from the target domain. The first transfers contextualised supervisedly pretrained representations, while the second method transfers pretrained weights, used to initialise the target model's parameters. Second, we perform a series of analysis to spot the limits of the above-mentioned proposed methods. We find that even if the proposed transfer learning approach enhances the performance on social media domain, a hidden negative transfer may mitigate the final gain brought by transfer learning. In addition, an interpretive analysis of the pretrained model, show that pretrained neurons may be biased by what they have learned from the source domain, thus struggle with learning uncommon target-specific patterns. Third, stemming from our analysis, we propose a new adaptation scheme which augments the target model with normalised, weighted and randomly initialised neurons that beget a better adaptation while maintaining the valuable source knowledge. Finally, we propose a model, that in addition to the pre-learned knowledge from the high-resource source-domain, takes advantage of various supervised NLP tasks
Schaub, Léon-Paul. "Dimensions mémorielles de l'interaction écrite humain-machine ˸ une approche cognitive par les modèles mnémoniques pour la détection et la correction des incohérences du système dans les dialogues orientés-tâche." Electronic Thesis or Diss., université Paris-Saclay, 2022. http://www.theses.fr/2022UPASG023.
Full textIn this work, we are interested in the place of task-oriented dialogue systems in both automatic language processing and human-machine interaction. In particular, we focus on the difference in information processing and memory use, from one turn to the next, by humans and machines, during a written chat conversation. After having studied the mechanisms of memory retention and recall in humans during a dialogue, in particular during the accomplishment of a task, we hypothesize that one of the elements that may explain why the performance of machines remains below that of humans, is the ability to possess not only an image of the user, but also an image of oneself, explicitly summoned during the inferences linked to the continuation of the dialogue. This translates into the following three axes for the system. First, by the anticipation, at a given turn of speech, of the next turn of the user. Secondly, by the detection of an inconsistency in one's own utterance, facilitated, as we demonstrate, by the anticipation of the user's next turn as an additional cue. Finally, by predicting the number of remaining turns in the dialogue in order to have a better vision of the dialogue progression, taking into account the potential presence of an incoherence in one's own utterance, this is what we call the dual model of the system, which represents both the user and the image that the system sends to the user. To implement these features, we exploit end-to-end memory networks, a recurrent neural network model that has the specificity not only to handle long dialogue histories (such as an RNN or an LSTM) but also to create reflection jumps, allowing to filter the information contained in both the user's utterance and the dialogue history. In addition, these three reflection jumps serve as a "natural" attention mechanism for the memory network, similar to a transformer decoder. For our study, we enhance a type of memory network called WMM2Seq (sequence-based working memory network) by adding our three features. This model is inspired by cognitive models of memory, presenting the concepts of episodic memory, semantic memory and working memory. It performs well on dialogue response generation tasks on the DSTC2 (human-machine in the restaurant domain) and MultiWOZ (multi-domain created with Wizard of Oz) corpora; these are the corpora we use for our experiments. The three axes mentioned above bring two main contributions to the existing. Firstly, it adds complexity to the intelligence of the dialogue system by providing it with a safeguard (detected inconsistencies). Second, it optimizes both the processing of information in the dialogue (more accurate or richer answers) and the duration of the dialogue. We evaluate the performance of our system with firstly the F1 score for the entities detected in each speech turn, secondly the BLEU score for the fluency of the system utterance and thirdly the joint accuracy for the success of the dialogue. The results obtained show that it would be interesting to direct research towards more cognitive models of memory management in order to reduce the performance gap in a human-machine dialogue
Coavoux, Maximin. "Discontinuous constituency parsing of morphologically rich languages." Thesis, Sorbonne Paris Cité, 2017. http://www.theses.fr/2017USPCC032.
Full textSyntactic parsing consists in assigning syntactic trees to sentences in natural language. Syntactic parsing of non-configurational languages, or languages with a rich inflectional morphology, raises specific problems. These languages suffer more from lexical data sparsity and exhibit word order variation phenomena more frequently. For these languages, exploiting information about the internal structure of word forms is crucial for accurate parsing. This dissertation investigates transition-based methods for robust discontinuous constituency parsing. First of all, we propose a multitask learning neural architecture that performs joint parsing and morphological analysis. Then, we introduce a new transition system that is able to predict discontinuous constituency trees, i.e.\ syntactic structures that can be seen as derivations of mildly context-sensitive grammars, such as LCFRS. Finally, we investigate the question of lexicalization in syntactic parsing. Some syntactic parsers are based on the hypothesis that constituent are organized around a lexical head and that modelling bilexical dependencies is essential to solve ambiguities. We introduce an unlexicalized transition system for discontinuous constituency parsing and a scoring model based on constituent boundaries. The resulting parser is simpler than lexicalized parser and achieves better results in both discontinuous and projective constituency parsing
Belissen, Valentin. "From Sign Recognition to Automatic Sign Language Understanding : Addressing the Non-Conventionalized Units." Electronic Thesis or Diss., université Paris-Saclay, 2020. http://www.theses.fr/2020UPASG064.
Full textSign Languages (SLs) have developed naturally in Deaf communities. With no written form, they are oral languages, using the gestural channel for expression and the visual channel for reception. These poorly endowed languages do not meet with a broad consensus at the linguistic level. These languages make use of lexical signs, i.e. conventionalized units of language whose form is supposed to be arbitrary, but also - and unlike vocal languages, if we don't take into account the co-verbal gestures - iconic structures, using space to organize discourse. Iconicity, which is defined as the existence of a similarity between the form of a sign and the meaning it carries, is indeed used at several levels of SL discourse.Most research in automatic Sign Language Recognition (SLR) has in fact focused on recognizing lexical signs, at first in the isolated case and then within continuous SL. The video corpora associated with such research are often relatively artificial, consisting of the repetition of elicited utterances in written form. Other corpora consist of interpreted SL, which may also differ significantly from natural SL, as it is strongly influenced by the surrounding vocal language.In this thesis, we wish to show the limits of this approach, by broadening this perspective to consider the recognition of elements used for the construction of discourse or within illustrative structures.To do so, we show the interest and the limits of the corpora developed by linguists. In these corpora, the language is natural and the annotations are sometimes detailed, but not always usable as input data for machine learning systems, as they are not necessarily complete or coherent. We then propose the redesign of a French Sign Language dialogue corpus, Dicta-Sign-LSF-v2, with rich and consistent annotations, following an annotation scheme shared by many linguists.We then propose a redefinition of the problem of automatic SLR, consisting in the recognition of various linguistic descriptors, rather than focusing on lexical signs only. At the same time, we discuss adapted metrics for relevant performance assessment.In order to perform a first experiment on the recognition of linguistic descriptors that are not only lexical, we then develop a compact and generalizable representation of signers in videos. This is done by parallel processing of the hands, face and upper body, using existing tools and models that we have set up. Besides, we preprocess these parallel representations to obtain a relevant feature vector. We then present an adapted and modular architecture for automatic learning of linguistic descriptors, consisting of a recurrent and convolutional neural network.Finally, we show through a quantitative and qualitative analysis the effectiveness of the proposed model, tested on Dicta-Sign-LSF-v2. We first carry out an in-depth analysis of the parameterization, evaluating both the learning model and the signer representation. The study of the model predictions then demonstrates the merits of the proposed approach, with a very interesting performance for the continuous recognition of four linguistic descriptors, especially in view of the uncertainty related to the annotations themselves. The segmentation of the latter is indeed subjective, and the very relevance of the categories used is not strongly demonstrated. Indirectly, the proposed model could therefore make it possible to measure the validity of these categories. With several areas for improvement being considered, particularly in terms of signer representation and the use of larger corpora, the results are very encouraging and pave the way for a wider understanding of continuous Sign Language Recognition
Djioua, Brahim. "Modélisation informatique d'une base de connaissances lexicales (DISSC) : réseaux polysémiques et schémas sémantico-cognitifs." Paris 4, 2000. http://www.theses.fr/2000PA040180.
Full textLinhares, Pontes Elvys. "Compressive Cross-Language Text Summarization." Thesis, Avignon, 2018. http://www.theses.fr/2018AVIG0232/document.
Full textThe popularization of social networks and digital documents increased quickly the informationavailable on the Internet. However, this huge amount of data cannot be analyzedmanually. Natural Language Processing (NLP) analyzes the interactions betweencomputers and human languages in order to process and to analyze natural languagedata. NLP techniques incorporate a variety of methods, including linguistics, semanticsand statistics to extract entities, relationships and understand a document. Amongseveral NLP applications, we are interested, in this thesis, in the cross-language textsummarization which produces a summary in a language different from the languageof the source documents. We also analyzed other NLP tasks (word encoding representation,semantic similarity, sentence and multi-sentence compression) to generate morestable and informative cross-lingual summaries.Most of NLP applications (including all types of text summarization) use a kind ofsimilarity measure to analyze and to compare the meaning of words, chunks, sentencesand texts in their approaches. A way to analyze this similarity is to generate a representationfor these sentences that contains the meaning of them. The meaning of sentencesis defined by several elements, such as the context of words and expressions, the orderof words and the previous information. Simple metrics, such as cosine metric andEuclidean distance, provide a measure of similarity between two sentences; however,they do not analyze the order of words or multi-words. Analyzing these problems,we propose a neural network model that combines recurrent and convolutional neuralnetworks to estimate the semantic similarity of a pair of sentences (or texts) based onthe local and general contexts of words. Our model predicted better similarity scoresthan baselines by analyzing better the local and the general meanings of words andmulti-word expressions.In order to remove redundancies and non-relevant information of similar sentences,we propose a multi-sentence compression method that compresses similar sentencesby fusing them in correct and short compressions that contain the main information ofthese similar sentences. We model clusters of similar sentences as word graphs. Then,we apply an integer linear programming model that guides the compression of theseclusters based on a list of keywords. We look for a path in the word graph that has goodcohesion and contains the maximum of keywords. Our approach outperformed baselinesby generating more informative and correct compressions for French, Portugueseand Spanish languages. Finally, we combine these previous methods to build a cross-language text summarizationsystem. Our system is an {English, French, Portuguese, Spanish}-to-{English,French} cross-language text summarization framework that analyzes the informationin both languages to identify the most relevant sentences. Inspired by the compressivetext summarization methods in monolingual analysis, we adapt our multi-sentencecompression method for this problem to just keep the main information. Our systemproves to be a good alternative to compress redundant information and to preserve relevantinformation. Our system improves informativeness scores without losing grammaticalquality for French-to-English cross-lingual summaries. Analyzing {English,French, Portuguese, Spanish}-to-{English, French} cross-lingual summaries, our systemsignificantly outperforms extractive baselines in the state of the art for all these languages.In addition, we analyze the cross-language text summarization of transcriptdocuments. Our approach achieved better and more stable scores even for these documentsthat have grammatical errors and missing information
Tian, Tian. "Domain Adaptation and Model Combination for the Annotation of Multi-source, Multi-domain Texts." Thesis, Paris 3, 2019. http://www.theses.fr/2019PA030003.
Full textThe increasing mass of User-Generated Content (UGC) on the Internet means that people are now willing to comment, edit or share their opinions on different topics. This content is now the main ressource for sentiment analysis on the Internet. Due to abbreviations, noise, spelling errors and all other problems with UGC, traditional Natural Language Processing (NLP) tools, including Named Entity Recognizers and part-of-speech (POS) taggers, perform poorly when compared to their usual results on canonical text (Ritter et al., 2011).This thesis deals with Named Entity Recognition (NER) on some User-Generated Content (UGC). We have created an evaluation dataset including multi-domain and multi-sources texts. We then developed a Conditional Random Fields (CRFs) model trained on User-Generated Content (UGC).In order to improve NER results in this context, we first developed a POStagger on UGC and used the predicted POS tags as a feature in the CRFs model. To turn UGC into canonical text, we also developed a normalization model using neural networks to propose a correct form for Non-Standard Words (NSW) in the UGC
各种社交网络应用使得互联网用户对各种话题的实时评价,编辑和分享成为可能。这类用户生成的文本内容(User Generated content)已成为社交网络上意见分析的主要目标和来源。但是,此类文本内容中包含的缩写,噪声(不规则词),拼写错误以及其他各种问题导致包括命名实体识别,词性标注在内的传统的自然语言处理工具的性能,相比良好组成的文本降低了许多【参见Ritter 2011】。本论文的主要目标是针对社交网络上用户生成文本内容的命名实体识别。我们首先建立了一个包含多来源,多领域文本的有标注的语料库作为标准评价语料库。然后,我们开发了一个由社交网络用户生成文本训练的基于条件随机场(Conditional Random Fields)的序列标注模型。基于改善这个命名实体识别模型的目的,我们又开发了另一个同样由社交网络用户生成内容训练的词性标注模型,并使用此模型预测的词性作为命名实体识别的条件随机场模型的特征。最后,为了将用户生成文本内容转换成相对标准的良好文本内容,我们开发了一个基于神经网络的词汇标准化模型,用以改正用户生成文本内容中的不标准字,并使用模型提供的改正形式作为命名实体识别的条件随机场模型的特征,借以改善原模型的性能。
Tourille, Julien. "Extracting Clinical Event Timelines : Temporal Information Extraction and Coreference Resolution in Electronic Health Records." Thesis, Université Paris-Saclay (ComUE), 2018. http://www.theses.fr/2018SACLS603/document.
Full textImportant information for public health is contained within Electronic Health Records (EHRs). The vast majority of clinical data available in these records takes the form of narratives written in natural language. Although free text is convenient to describe complex medical concepts, it is difficult to use for medical decision support, clinical research or statistical analysis.Among all the clinical aspects that are of interest in these records, the patient timeline is one of the most important. Being able to retrieve clinical timelines would allow for a better understanding of some clinical phenomena such as disease progression and longitudinal effects of medications. It would also allow to improve medical question answering and clinical outcome prediction systems. Accessing the clinical timeline is needed to evaluate the quality of the healthcare pathway by comparing it to clinical guidelines, and to highlight the steps of the pathway where specific care should be provided.In this thesis, we focus on building such timelines by addressing two related natural language processing topics which are temporal information extraction and clinical event coreference resolution.Our main contributions include a generic feature-based approach for temporal relation extraction that can be applied to documents written in English and in French. We devise a neural based approach for temporal information extraction which includes categorical features.We present a neural entity-based approach for coreference resolution in clinical narratives. We perform an empirical study to evaluate how categorical features and neural network components such as attention mechanisms and token character-level representations influence the performance of our coreference resolution approach
Colin, Émilie. "Traitement automatique des langues et génération automatique d'exercices de grammaire." Electronic Thesis or Diss., Université de Lorraine, 2020. http://www.theses.fr/2020LORR0059.
Full textOur perspectives are educational, to create grammar exercises for French. Paraphrasing is an operation of reformulation. Our work tends to attest that sequence-to-sequence models are not simple repeaters but can learn syntax. First, by combining various models, we have shown that the representation of information in multiple forms (using formal data (RDF), coupled with text to extend or reduce it, or only text) allows us to exploit a corpus from different angles, increasing the diversity of outputs, exploiting the syntactic levers put in place. We also addressed a recurrent problem, that of data quality, and obtained paraphrases with a high syntactic adequacy (up to 98% coverage of the demand) and a very good linguistic level. We obtain up to 83.97 points of BLEU-4*, 78.41 more than our baseline average, without syntax leverage. This rate indicates a better control of the outputs, which are varied and of good quality in the absence of syntax leverage. Our idea was to be able to work from raw text : to produce a representation of its meaning. The transition to French text was also an imperative for us. Working from plain text, by automating the procedures, allowed us to create a corpus of more than 450,000 sentence/representation pairs, thanks to which we learned to generate massively correct texts (92% on qualitative validation). Anonymizing everything that is not functional contributed significantly to the quality of the results (68.31 of BLEU, i.e. +3.96 compared to the baseline, which was the generation of text from non-anonymized data). This second work can be applied the integration of a syntax lever guiding the outputs. What was our baseline at time 1 (generate without constraint) would then be combined with a constrained model. By applying an error search, this would allow the constitution of a silver base associating representations to texts. This base could then be multiplied by a reapplication of a generation under constraint, and thus achieve the applied objective of the thesis. The formal representation of information in a language-specific framework is a challenging task. This thesis offers some ideas on how to automate this operation. Moreover, we were only able to process relatively short sentences. The use of more recent neural modelswould likely improve the results. The use of appropriate output strokes would allow for extensive checks. *BLEU : quality of a text (scale from 0 (worst) to 100 (best), Papineni et al. (2002))
Caglayan, Ozan. "Multimodal Machine Translation." Thesis, Le Mans, 2019. http://www.theses.fr/2019LEMA1016/document.
Full textMachine translation aims at automatically translating documents from one language to another without human intervention. With the advent of deep neural networks (DNN), neural approaches to machine translation started to dominate the field, reaching state-ofthe-art performance in many languages. Neural machine translation (NMT) also revived the interest in interlingual machine translation due to how it naturally fits the task into an encoder-decoder framework which produces a translation by decoding a latent source representation. Combined with the architectural flexibility of DNNs, this framework paved the way for further research in multimodality with the objective of augmenting the latent representations with other modalities such as vision or speech, for example. This thesis focuses on a multimodal machine translation (MMT) framework that integrates a secondary visual modality to achieve better and visually grounded language understanding. I specifically worked with a dataset containing images and their translated descriptions, where visual context can be useful forword sense disambiguation, missing word imputation, or gender marking when translating from a language with gender-neutral nouns to one with grammatical gender system as is the case with English to French. I propose two main approaches to integrate the visual modality: (i) a multimodal attention mechanism that learns to take into account both sentence and convolutional visual representations, (ii) a method that uses global visual feature vectors to prime the sentence encoders and the decoders. Through automatic and human evaluation conducted on multiple language pairs, the proposed approaches were demonstrated to be beneficial. Finally, I further show that by systematically removing certain linguistic information from the input sentences, the true strength of both methods emerges as they successfully impute missing nouns, colors and can even translate when parts of the source sentences are completely removed
Enguehard, Chantal. "Acquisition naturelle automatique d'un réseau sémantique." Compiègne, 1992. http://www.theses.fr/1992COMPD527.
Full textPoitevin, Christine. "Contribution à une définition du domaine des industries de la langue : l'élucidation des statuts juridiques." Paris 8, 1993. http://www.theses.fr/1993PA080861.
Full textThe automatic processing of the natural language spoken by humans has gradually become a field of industrial implementation. The concept of language industries indicates a set of activities aiming and having the natural language spoken or written by humans handled, interpreted or engendered by machines. These activities spread from research to trade going through the industrial development of products. But the necessary investements to the achievement of products in the field of language industries cannot be granted in a judicial environment that would not bring a satisfactory level of security. Therefore, law ought to provide a relevant response to the need of companies to be protected and should fulfil its regulating role. The products created by the language industries are unsubstantial goods and can benefit from the protection of copyright. These products are also complex computer creations that are made up of two distinct parts: a software part and a dictionary part. Each of its components is liable to be protected by a different right (the law of marche 17th 1957 for the dictionary part and the law of july 3rd 1985 for the software part). The response to the problem of the judicial status can only be brought after wondering if the products of the field language industries considered as a whole can be qualified as software
Gruer, Juan Pablo. "Eléments de synchronisation pour un langage temps-réel de commande de procédés." Mulhouse, 1989. http://www.theses.fr/1989MULH0105.
Full textMarzinotto, Gabriel. "Semantic frame based analysis using machine learning techniques : improving the cross-domain generalization of semantic parsers." Electronic Thesis or Diss., Aix-Marseille, 2019. http://www.theses.fr/2019AIXM0483.
Full textMaking semantic parsers robust to lexical and stylistic variations is a real challenge with many industrial applications. Nowadays, semantic parsing requires the usage of domain-specific training corpora to ensure acceptable performances on a given domain. Transfer learning techniques are widely studied and adopted when addressing this lack of robustness, and the most common strategy is the usage of pre-trained word representations. However, the best parsers still show significant performance degradation under domain shift, evidencing the need for supplementary transfer learning strategies to achieve robustness. This work proposes a new benchmark to study the domain dependence problem in semantic parsing. We use this bench to evaluate classical transfer learning techniques and to propose and evaluate new techniques based on adversarial learning. All these techniques are tested on state-of-the-art semantic parsers. We claim that adversarial learning approaches can improve the generalization capacities of models. We test this hypothesis on different semantic representation schemes, languages and corpora, providing experimental results to support our hypothesis
Soriano-Morales, Edmundo-Pavel. "Hypergraphs and information fusion for term representation enrichment : applications to named entity recognition and word sense disambiguation." Thesis, Lyon, 2018. http://www.theses.fr/2018LYSE2009/document.
Full textMaking sense of textual data is an essential requirement in order to make computers understand our language. To extract actionable information from text, we need to represent it by means of descriptors before using knowledge discovery techniques.The goal of this thesis is to shed light into heterogeneous representations of words and how to leverage them while addressing their implicit sparse nature.First, we propose a hypergraph network model that holds heterogeneous linguistic data in a single unified model. In other words, we introduce a model that represents words by means of different linguistic properties and links them together accordingto said properties. Our proposition differs to other types of linguistic networks in that we aim to provide a general structure that can hold several types of descriptive text features, instead of a single one as in most representations. This representationmay be used to analyze the inherent properties of language from different points of view, or to be the departing point of an applied NLP task pipeline. Secondly, we employ feature fusion techniques to provide a final single enriched representation that exploits the heterogeneous nature of the model and alleviates the sparseness of each representation.These types of techniques are regularly used exclusively to combine multimedia data. In our approach, we consider different text representations as distinct sources of information which can be enriched by themselves. This approach has not been explored before, to the best of our knowledge. Thirdly, we propose an algorithm that exploits the characteristics of the network to identify and group semantically related words by exploiting the real-world properties of the networks. In contrast with similar methods that are also based on the structure of the network, our algorithm reduces the number of required parameters and more importantly, allows for the use of either lexical or syntactic networks to discover said groups of words, instead of the singletype of features usually employed.We focus on two different natural language processing tasks: Word Sense Induction and Disambiguation (WSI/WSD), and Named Entity Recognition (NER). In total, we test our propositions on four different open-access datasets. The results obtained allow us to show the pertinence of our contributions and also give us some insights into the properties of heterogeneous features and their combinations with fusion methods. Specifically, our experiments are twofold: first, we show that using fusion-enriched heterogeneous features, coming from our proposed linguistic network, we outperform the performance of single features’ systems and other basic baselines. We note that using single fusion operators is not efficient compared to using a combination of them in order to obtain a final space representation. We show that the features added by each combined fusion operation are important towards the models predicting the appropriate classes. We test the enriched representations on both WSI/WSD and NER tasks. Secondly, we address the WSI/WSD task with our network-based proposed method. While based on previous work, we improve it by obtaining better overall performance and reducing the number of parameters needed. We also discuss the use of either lexical or syntactic networks to solve the task.Finally, we parse a corpus based on the English Wikipedia and then store it following the proposed network model. The parsed Wikipedia version serves as a linguistic resource to be used by other researchers. Contrary to other similar resources, insteadof just storing its part of speech tag and its dependency relations, we also take into account the constituency-tree information of each word analyzed. The hope is for this resource to be used on future developments without the need to compile suchresource from zero
Dupont, Yoann. "La structuration dans les entités nommées." Thesis, Sorbonne Paris Cité, 2017. http://www.theses.fr/2017USPCA100/document.
Full textNamed entity recognition is a crucial discipline of NLP. It is used to extract relations between named entities, which allows the construction of knowledge bases (Surdeanu and Ji, 2014), automatic summary (Nobata et al., 2002) and so on. Our interest in this thesis revolves around structuration phenomena that surround them.We distinguish here two kinds of structural elements in named entities. The first one are recurrent substrings, that we will call the caracteristic affixes of a named entity. The second type of element is tokens with a good discriminative power, which we call trigger tokens of named entities. We will explain here the algorithm we provided to extract such affixes, which we will compare to Morfessor (Creutz and Lagus, 2005b). We will then apply the same algorithm to extract trigger tokens, which we will use for French named entity recognition and postal address extraction.Another form of structuration for named entities is of a syntactic nature. It follows an overlapping or tree structure. We propose a novel kind of linear tagger cascade which have not been used before for structured named entity recognition, generalising other previous methods that are only able to recognise named entities of a fixed depth or being unable to model certain characteristics of the structure. Ours, however, can do both.Throughout this thesis, we compare two machine learning methods, CRFs and neural networks, for which we will compare respective advantages and drawbacks
Nana, jipmo Coriane. "Intégration du web social dans les systèmes de recommandation." Thesis, Université Paris-Saclay (ComUE), 2017. http://www.theses.fr/2017SACLC082/document.
Full textThe social Web grows more and more and gives through the web, access to a wide variety of resources, like sharing sites such as del.icio.us, exchange messages as Twitter, or social networks with the professional purpose such as LinkedIn, or more generally for social purposes, such as Facebook and LiveJournal. The same individual can be registered and active on different social networks (potentially having different purposes), in which it publishes various information, which are constantly growing, such as its name, locality, communities, various activities. The information (textual), given the international dimension of the Web, is inherently multilingual and intrinsically ambiguous, since it is published in natural language in a free vocabulary by individuals from different origin. They are also important, specially for applications seeking to know their users in order to better understand their needs, activities and interests. The objective of our research is to exploit using essentially the Wikpédia encyclopedia, the textual resources extracted from the different social networks of the same individual in order to construct his characterizing profile, which can be exploited in particular by applications seeking to understand their users, such as recommendation systems. In particular, we conducted a study to characterize the personality traits of users. Many experiments, analyzes and evaluations were carried out on real data collected from different social networks
Lumbreras, Alberto. "Automatic role detection in online forums." Thesis, Lyon, 2016. http://www.theses.fr/2016LYSE2111/document.
Full textThis thesis addresses the problem of detecting user roles in online discussion forums. A role may be defined as the set of behaviors characteristic of a person or a position. In discussion forums, behaviors are primarily observed through conversations. Hence, we focus our attention on how users discuss. We propose three methods to detect groups of users with similar conversational behaviors.Our first method for the detection of roles is based on conversational structures. Weapply different notions of neighborhood for posts in tree graphs (radius-based, order-based, and time-based) and compare the conversational patterns that they detect as well as the clusters of users with similar conversational patterns.Our second method is based on stochastic models of growth for conversation threads.Building upon these models we propose a method to find groups of users that tend to reply to the same type of posts. We show that, while there are clusters of users with similar replying patterns, there is no strong evidence that these behaviors are predictive of future behaviors |except for some groups of users with extreme behaviors.In out last method, we integrate the type of data used in the two previous methods(feature-based and behavioral or functional-based) and show that we can find clusters using fewer examples. The model exploits the idea that users with similar features have similar behaviors
Parmentier, François. "Spécification d'une architecture émergente fondée sur le raisonnement par analogie : application aux références bibliographiques." Phd thesis, Université Henri Poincaré - Nancy I, 1998. http://tel.archives-ouvertes.fr/tel-00003024.
Full textcréativité il adapte son comportement en fonction de la solution courante. Nous l'avons appliqué à la reconnaissance automatique de la structure logique (des champs) de références bibliographiques dans les articles scientifiques (en format uniquement physique, c'est-à-dire en PostScript). Le modèle, appelé Réseau de Concepts, s'apparentant à la fois aux réseaux sémantiques et aux réseaux de neurones, est construit automatiquement à partir d'une base de références BIBTeX. Le système utilise les co-occurrences entre les termes des références pour rapprocher dans le modèle ceux qui sont conceptuellement voisins. Le principe de l'analogie est utilisé sur les références de la base : quand le système rencontre une référence inconnue, il fait l'analogie avec la partie physique de la base et essaye de proposer une solution correspondante. Les résultats obtenus, bien que modérés (65,5% de reconnaissance), laissent augurer des résultats encore meilleurs, après optimisation du système.
Chatelain, Clément. "Extraction de séquences numériques dans des documents manuscrits quelconques." Phd thesis, Université de Rouen, 2006. http://tel.archives-ouvertes.fr/tel-00143090.
Full textChatelain, Clément. "Extraction de séquences numériques dans des documents manuscrits quelconques." Phd thesis, Rouen, 2006. http://www.theses.fr/2006ROUES056.
Full textWithin the framework of the automatic processing of incoming mail documents, we present in this thesis the conception and development of a numerical field extraction system in weakly constrained handwritten documents. Although the recognition of isolated handwritten entities can be considered as a partially solved problem, the extraction of information in images of complex and free-layout documents is still a challenge. This problem requires the implementation of both handwriting recognition and information extraction methods inspired by approaches developed within the field of information extraction in electronic documents. Our contribution consists in the conception and the implementation of two different strategies: the first extends classical handwriting recognition methods, while the second is inspired from approaches used within the field of information extraction in electronic documents. The results obtained on a real handwritten mail database show that our second approach is significantly better. Finally, a complete, generic and efficient system is produced, answering one of the emergent perspectives in the field of the automatic reading of handwritten documents: the extraction of complex information in images of documents
Vukotic, Verdran. "Deep Neural Architectures for Automatic Representation Learning from Multimedia Multimodal Data." Thesis, Rennes, INSA, 2017. http://www.theses.fr/2017ISAR0015/document.
Full textIn this dissertation, the thesis that deep neural networks are suited for analysis of visual, textual and fused visual and textual content is discussed. This work evaluates the ability of deep neural networks to learn automatic multimodal representations in either unsupervised or supervised manners and brings the following main contributions:1) Recurrent neural networks for spoken language understanding (slot filling): different architectures are compared for this task with the aim of modeling both the input context and output label dependencies.2) Action prediction from single images: we propose an architecture that allow us to predict human actions from a single image. The architecture is evaluated on videos, by utilizing solely one frame as input.3) Bidirectional multimodal encoders: the main contribution of this thesis consists of neural architecture that translates from one modality to the other and conversely and offers and improved multimodal representation space where the initially disjoint representations can translated and fused. This enables for improved multimodal fusion of multiple modalities. The architecture was extensively studied an evaluated in international benchmarks within the task of video hyperlinking where it defined the state of the art today.4) Generative adversarial networks for multimodal fusion: continuing on the topic of multimodal fusion, we evaluate the possibility of using conditional generative adversarial networks to lean multimodal representations in addition to providing multimodal representations, generative adversarial networks permit to visualize the learned model directly in the image domain
Zhang, Saizheng. "Recurrent neural models and related problems in natural language processing." Thèse, 2019. http://hdl.handle.net/1866/22663.
Full textLin, Zhouhan. "Deep neural networks for natural language processing and its acceleration." Thèse, 2019. http://hdl.handle.net/1866/23438.
Full textThis thesis by article consists of four articles which contribute to the field of deep learning, specifically in the acceleration of training through low-precision networks, and the application of deep neural networks on natural language processing. In the first article, we investigate a neural network training scheme that eliminates most of the floating-point multiplications. This approach consists of binarizing or ternarizing the weights in the forward propagation and quantizing the hidden states in the backward propagation, which converts multiplications to sign changes and binary shifts. Experimental results on datasets from small to medium size show that this approach result in even better performance than standard stochastic gradient descent training, paving the way to fast, hardware-friendly training of neural networks. In the second article, we proposed a structured self-attentive sentence embedding that extracts interpretable sentence representations in matrix form. We demonstrate improvements on 3 different tasks: author profiling, sentiment classification and textual entailment. Experimental results show that our model yields a significant performance gain compared to other sentence embedding methods in all of the 3 tasks. In the third article, we propose a hierarchical model with dynamical computation graph for sequential data that learns to construct a tree while reading the sequence. The model learns to create adaptive skip-connections that ease the learning of long-term dependencies through constructing recurrent cells in a recursive manner. The training of the network can either be supervised training by giving golden tree structures, or through reinforcement learning. We provide preliminary experiments in 3 different tasks: a novel Math Expression Evaluation (MEE) task, a well-known propositional logic task, and language modelling tasks. Experimental results show the potential of the proposed approach. In the fourth article, we propose a novel constituency parsing method with neural networks. The model predicts the parse tree structure by predicting a real valued scalar, named syntactic distance, for each split position in the input sentence. The order of the relative values of these syntactic distances then determine the parse tree structure by specifying the order in which the split points will be selected, recursively partitioning the input, in a top-down fashion. Our proposed approach was demonstrated with competitive performance on Penn Treebank dataset, and the state-of-the-art performance on Chinese Treebank dataset.