La nature diachronique des bulletins d'information provoque de fortes variations du contenu linguistique et du vocabulaire dans ces documents. Dans le cadre de la reconnaissance automatique de la parole, cela conduit au problème de mots hors vocabulaire (Out-Of-Vocabulary, OOV). La plupart des mots OOV sont des noms propres. Les noms propres sont très importants pour l'indexation automatique de contenus audio-vidéo. De plus, leur bonne identification est importante pour des transcriptions automatiques fiables. Le but de cette thèse est de proposer des méthodes pour récupérer les noms propres manquants dans un système de reconnaissance. Nous proposons de modéliser le contexte sémantique et d'utiliser des informations thématiques contenus dans les documents audio à transcrire. Des modèles probabilistes de thème et des projections dans un espace continu obtenues à l'aide de réseaux de neurones sont explorés pour la tâche de récupération des noms propres pertinents. Une évaluation approfondie de ces représentations contextuelles a été réalisée. Pour modéliser le contexte de nouveaux mots plus efficacement, nous proposons des réseaux de neurones qui maximisent la récupération des noms propres pertinents. En s'appuyant sur ce modèle, nous proposons un nouveau modèle (Neural Bag-of-Weighted-Words, NBOW2) qui permet d'estimer un degré d'importance pour chacun des mots du document et a la capacité de capturer des mots spécifiques à ce document. Des expériences de reconnaissance automatique de bulletins d'information télévisés montrent l'efficacité du modèle proposé. L'évaluation de NBOW2 sur d'autres tâches telles que la classification de textes montre des bonnes performances / The diachronic nature of broadcast news causes frequent variations in the linguistic content and vocabulary, leading to the problem of Out-Of-Vocabulary (OOV) words in automatic speech recognition. Most of the OOV words are found to be proper names whereas proper names are important for automatic indexing of audio-video content as well as for obtaining reliable automatic transcriptions. The goal of this thesis is to model the semantic and topical context of new proper names in order to retrieve those which are relevant to the spoken content in the audio document. Training context models is a challenging problem in this task because several new names come with a low amount of data and the context model should be robust to errors in the automatic transcription. Probabilistic topic models and word embeddings from neural network models are explored for the task of retrieval of relevant proper names. A thorough evaluation of these contextual representations is performed. It is argued that these representations, which are learned in an unsupervised manner, are not the best for the given retrieval task. Neural network context models trained with an objective to maximise the retrieval performance are proposed. The proposed Neural Bag-of-Weighted-Words (NBOW2) model learns to assign a degree of importance to input words and has the ability to capture task specific key-words. Experiments on automatic speech recognition on French broadcast news videos demonstrate the effectiveness of the proposed models. Evaluation of the NBOW2 model on standard text classification tasks shows that it learns interesting information and gives best classification accuracies among the BOW models

An ontology is a knowledge-representation structure, where words, terms or concepts are defined by their mutual hierarchical relations. Ontologies are becoming ever more prevalent in the world of natural language processing, where we currently see a tendency towards using semantics for solving a variety of tasks, particularly tasks related to information access. Ontologies, taxonomies and thesauri (all related notions) are also used in various variants by humans, to standardize business transactions or for finding conceptual relations between terms in, e.g., the medical domain. The acquisition of machine-readable, domain-specific semantic knowledge is time consuming and prone to inconsistencies. The field of ontology learning therefore provides tools for automating the construction of domain ontologies (ontologies describing the entities and relations within a particular field of interest), by analyzing large quantities of domain-specific texts. This thesis studies three main topics within the field of ontology learning. First, we examine which sources of information are useful within an ontology learning system and how the information sources can be combined effectively. Secondly, we do this with a special focus on cross-language text collections, to see if we can learn more from studying several languages at once, than we can from a single-language text collection. Finally, we investigate new approaches to formal and automatic evaluation of the quality of a learned ontology. We demonstrate how to combine information sources from different languages and use them to train automatic classifiers to recognize lexico-semantic relations. The cross-language data is shown to have a positive effect on the quality of the learned ontologies. We also give theoretical and experimental results, showing that our ontology evaluation method is a good complement to and in some aspects improves on the evaluation measures in use today. / För att köpa boken skicka en beställning till exp@ling.su.se/ To order the book send an e-mail to exp@ling.su.se

Les systèmes de traitement automatique des langues reposent souvent sur l'idée que le langage est compositionnel, c'est-à-dire que le sens d'une entité linguistique peut être déduite à partir du sens de ses parties. Cette supposition ne s’avère pas vraie dans le cas des expressions polylexicales (EPLs). Par exemple, une "poule mouillée" n'est ni une poule, ni nécessairement mouillée. Les techniques pour déduire le sens des mots en fonction de leur distribution dans le texte ont obtenu de bons résultats sur plusieurs tâches, en particulier depuis l'apparition des word embeddings. Cependant, la représentation des EPLs reste toujours un problème non résolu. En particulier, on ne sait pas comment prédire avec précision, à partir des corpus, si une EPL donnée doit être traitée comme une unité indivisible (p.ex. "carton plein") ou comme une combinaison du sens de ses parties (p.ex. "eau potable"). Cette thèse propose un cadre méthodologique pour la prédiction de compositionnalité d'EPLs fondé sur des représentations de la sémantique distributionnelle, que nous instancions à partir d’une variété de paramètres. Nous présenterons une évaluation complète de l'impact de ces paramètres sur trois nouveaux ensembles de données modélisant la compositionnalité d'EPLs, en anglais, français et portugais. Finalement, nous présenterons une évaluation extrinsèque des niveaux de compositionnalité prédits par le modèle dans le contexte d’un système d'identification d'EPLs. Les résultats suggèrent que le choix spécifique de modèle distributionnel et de paramètres de corpus peut produire des prédictions de compositionnalité qui sont comparables à celles présentées dans l'état de l'art. / Natural language processing systems often rely on the idea that language is compositional, that is, the meaning of a linguistic entity can be inferred from the meaning of its parts. This expectation fails in the case of multiword expressions (MWEs). For example, a person who is a "sitting duck" is neither a duck nor necessarily sitting. Modern computational techniques for inferring word meaning based on the distribution of words in the text have been quite successful at multiple tasks, especially since the rise of word embedding approaches. However, the representation of MWEs still remains an open problem in the field. In particular, it is unclear how one could predict from corpora whether a given MWE should be treated as an indivisible unit (e.g. "nut case") or as some combination of the meaning of its parts (e.g. "engine room"). This thesis proposes a framework of MWE compositionality prediction based on representations of distributional semantics, which we instantiate under a variety of parameters. We present a thorough evaluation of the impact of these parameters on three new datasets of MWE compositionality, encompassing English, French and Portuguese MWEs. Finally, we present an extrinsic evaluation of the predicted levels of MWE compositionality on the task of MWE identification. Our results suggest that the proper choice of distributional model and corpus parameters can produce compositionality predictions that are comparable to the state of the art.

This thesis describes the work of creating an automatic system for identifying synonyms and semantically related words in medical texts. Before this work, as a part of the project E-care@home, medical texts have been classified as either lay or specialized by both a lay annotator and an expert annotator. The lay annotator, in this case, is a person without any medical knowledge, whereas the expert annotator has professional knowledge in medicine. Using these texts made it possible to create co-occurrences matrices from which the related words could be identified. Fifteen medical terms were chosen as system input. The Dice similarity of these words in a context window of ten words around them was calculated. As output, five candidate related terms for each medical term was returned. Only unigrams were considered. The candidate related terms were evaluated using a questionnaire, where 223 healthcare professionals rated the similarity using a scale from one to five. A Fleiss kappa test showed that the agreement among these raters was 0.28, which is a fair agreement. The evaluation further showed that there was a significant correlation between the human ratings and the relatedness score (Dice similarity). That is, words with higher Dice similarity tended to get a higher human rating. However, the Dice similarity interval in which the words got the highest average human rating was 0.35-0.39. This result means that there is much room for improving the system. Further developments of the system should remove the unigram limitation and expand the corpus the provide a more accurate and reliable result.

This thesis is about the problem of compositionality in distributional semantics. Distributional semantics presupposes that the meanings of words are a function of their occurrences in textual contexts. It models words as distributions over these contexts and represents them as vectors in high dimensional spaces. The problem of compositionality for such models concerns itself with how to produce distributional representations for larger units of text (such as a verb and its arguments) by composing the distributional representations of smaller units of text (such as individual words). This thesis focuses on a particular approach to this compositionality problem, namely using the categorical framework developed by Coecke, Sadrzadeh, and Clark, which combines syntactic analysis formalisms with distributional semantic representations of meaning to produce syntactically motivated composition operations. This thesis shows how this approach can be theoretically extended and practically implemented to produce concrete compositional distributional models of natural language semantics. It furthermore demonstrates that such models can perform on par with, or better than, other competing approaches in the field of natural language processing. There are three principal contributions to computational linguistics in this thesis. The first is to extend the DisCoCat framework on the syntactic front and semantic front, incorporating a number of syntactic analysis formalisms and providing learning procedures allowing for the generation of concrete compositional distributional models. The second contribution is to evaluate the models developed from the procedures presented here, showing that they outperform other compositional distributional models present in the literature. The third contribution is to show how using category theory to solve linguistic problems forms a sound basis for research, illustrated by examples of work on this topic, that also suggest directions for future research.

Distributional semantics allows models of linguistic meaning to be derived from observations of language use in large amounts of text. By modeling the meaning of words in semantic (vector) space on the basis of co-occurrence information, distributional semantics permits a quantitative interpretation of (relative) word meaning in an unsupervised setting, i.e., human annotations are not required. The ability to obtain inexpensive word representations in this manner helps to alleviate the bottleneck of fully supervised approaches to natural language processing, especially since models of distributional semantics are data-driven and hence agnostic to both language and domain. All that is required to obtain distributed word representations is a sizeable corpus; however, the composition of the semantic space is not only affected by the underlying data but also by certain model hyperparameters. While these can be optimized for a specific downstream task, there are currently limitations to the extent the many aspects of semantics can be captured in a single model. This dissertation investigates the possibility of capturing multiple aspects of lexical semantics by adopting the ensemble methodology within a distributional semantic framework to create ensembles of semantic spaces. To that end, various strategies for creating the constituent semantic spaces, as well as for combining them, are explored in a number of studies. The notion of semantic space ensembles is generalizable across languages and domains; however, the use of unsupervised methods is particularly valuable in low-resource settings, in particular when annotated corpora are scarce, as in the domain of Swedish healthcare. The semantic space ensembles are here empirically evaluated for tasks that have promising applications in healthcare. It is shown that semantic space ensembles – created by exploiting various corpora and data types, as well as by adjusting model hyperparameters such as the size of the context window and the strategy for handling word order within the context window – are able to outperform the use of any single constituent model on a range of tasks. The semantic space ensembles are used both directly for k-nearest neighbors retrieval and for semi-supervised machine learning. Applying semantic space ensembles to important medical problems facilitates the secondary use of healthcare data, which, despite its abundance and transformative potential, is grossly underutilized. / <p>At the time of the doctoral defense, the following papers were unpublished and had a status as follows: Paper 4 and 5: Unpublished conference papers.</p> / High-Performance Data Mining for Drug Effect Detection

abstract: In the current millennium, extensive use of computers and the internet caused an exponential increase in information. Few research areas are as important as information extraction, which primarily involves extracting concepts and the relations between them from free text. Limitations in the size of training data, lack of lexicons and lack of relationship patterns are major factors for poor performance in information extraction. This is because the training data cannot possibly contain all concepts and their synonyms; and it contains only limited examples of relationship patterns between concepts. Creating training data, lexicons and relationship patterns is expensive, especially in the biomedical domain (including clinical notes) because of the depth of domain knowledge required of the curators. Dictionary-based approaches for concept extraction in this domain are not sufficient to effectively overcome the complexities that arise because of the descriptive nature of human languages. For example, there is a relatively higher amount of abbreviations (not all of them present in lexicons) compared to everyday English text. Sometimes abbreviations are modifiers of an adjective (e.g. CD4-negative) rather than nouns (and hence, not usually considered named entities). There are many chemical names with numbers, commas, hyphens and parentheses (e.g. t(3;3)(q21;q26)), which will be separated by most tokenizers. In addition, partial words are used in place of full words (e.g. up- and downregulate); and some of the words used are highly specialized for the domain. Clinical notes contain peculiar drug names, anatomical nomenclature, other specialized names and phrases that are not standard in everyday English or in published articles (e.g. "l shoulder inj"). State of the art concept extraction systems use machine learning algorithms to overcome some of these challenges. However, they need a large annotated corpus for every concept class that needs to be extracted. A novel natural language processing approach to minimize this limitation in concept extraction is proposed here using distributional semantics. Distributional semantics is an emerging field arising from the notion that the meaning or semantics of a piece of text (discourse) depends on the distribution of the elements of that discourse in relation to its surroundings. Distributional information from large unlabeled data is used to automatically create lexicons for the concepts to be tagged, clusters of contextually similar words, and thesauri of distributionally similar words. These automatically generated lexical resources are shown here to be more useful than manually created lexicons for extracting concepts from both literature and narratives. Further, machine learning features based on distributional semantics are shown to improve the accuracy of BANNER, and could be used in other machine learning systems such as cTakes to improve their performance. In addition, in order to simplify the sentence patterns and facilitate association extraction, a new algorithm using a "shotgun" approach is proposed. The goal of sentence simplification has traditionally been to reduce the grammatical complexity of sentences while retaining the relevant information content and meaning to enable better readability for humans and enhanced processing by parsers. Sentence simplification is shown here to improve the performance of association extraction systems for both biomedical literature and clinical notes. It helps improve the accuracy of protein-protein interaction extraction from the literature and also improves relationship extraction from clinical notes (such as between medical problems, tests and treatments). Overall, the two main contributions of this work include the application of sentence simplification to association extraction as described above, and the use of distributional semantics for concept extraction. The proposed work on concept extraction amalgamates for the first time two diverse research areas -distributional semantics and information extraction. This approach renders all the advantages offered in other semi-supervised machine learning systems, and, unlike other proposed semi-supervised approaches, it can be used on top of different basic frameworks and algorithms. / Dissertation/Thesis / Ph.D. Biomedical Informatics 2011

La compréhension du langage naturel repose souvent sur des raisonnements de sens commun, pour lesquels la connaissance de relations sémantiques, en particulier entre prédicats verbaux, peut être nécessaire. Cette thèse porte sur la problématique de l'utilisation d'une méthode distributionnelle pour extraire automatiquement les informations sémantiques nécessaires à ces inférences de sens commun. Des associations typiques entre des paires de prédicats et un ensemble de relations sémantiques (causales, temporelles, de similarité, d'opposition, partie/tout) sont extraites de grands corpus, par l'exploitation de la présence de connecteurs du discours signalant typiquement ces relations. Afin d'apprécier ces associations, nous proposons plusieurs mesures de signifiance inspirées de la littérature ainsi qu'une mesure novatrice conçue spécifiquement pour évaluer la force du lien entre les deux prédicats et la relation. La pertinence de ces mesures est évaluée par le calcul de leur corrélation avec des jugements humains, obtenus par l'annotation d'un échantillon de paires de verbes en contexte discursif. L'application de cette méthodologie sur des corpus de langue française et anglaise permet la construction d'une ressource disponible librement, Lecsie (Linked Events Collection for Semantic Information Extraction). Celle-ci est constituée de triplets: des paires de prédicats associés à une relation; à chaque triplet correspondent des scores de signifiance obtenus par nos mesures.Cette ressource permet de dériver des représentations vectorielles de paires de prédicats qui peuvent être utilisées comme traits lexico-sémantiques pour la construction de modèles pour des applications externes. Nous évaluons le potentiel de ces représentations pour plusieurs applications. Concernant l'analyse du discours, les tâches de la prédiction d'attachement entre unités du discours, ainsi que la prédiction des relations discursives spécifiques les reliant, sont explorées. En utilisant uniquement les traits provenant de notre ressource, nous obtenons des améliorations significatives pour les deux tâches, par rapport à plusieurs bases de référence, notamment des modèles utilisant d'autres types de représentations lexico-sémantiques. Nous proposons également de définir des ensembles optimaux de connecteurs mieux adaptés à des applications sur de grands corpus, en opérant une réduction de dimension dans l'espace des connecteurs, au lieu d'utiliser des groupes de connecteurs composés manuellement et correspondant à des relations prédéfinies. Une autre application prometteuse explorée dans cette thèse concerne les relations entre cadres sémantiques (semantic frames, e.g. FrameNet): la ressource peut être utilisée pour enrichir cette structure par des relations potentielles entre frames verbaux à partir des associations entre leurs verbes. Ces applications diverses démontrent les contributions prometteuses amenées par notre approche permettant l'extraction non supervisée de relations sémantiques. / Natural language understanding often relies on common-sense reasoning, for which knowledge about semantic relations, especially between verbal predicates, may be required. This thesis addresses the challenge of using a distibutional method to automatically extract the necessary semantic information for common-sense inference. Typical associations between pairs of predicates and a targeted set of semantic relations (causal, temporal, similarity, opposition, part/whole) are extracted from large corpora, by exploiting the presence of discourse connectives which typically signal these semantic relations. In order to appraise these associations, we provide several significance measures inspired from the literature as well as a novel measure specifically designed to evaluate the strength of the link between the two predicates and the relation. The relevance of these measures is evaluated by computing their correlations with human judgments, based on a sample of verb pairs annotated in context. The application of this methodology to French and English corpora leads to the construction of a freely available resource, Lecsie (Linked Events Collection for Semantic Information Extraction), which consists of triples: pairs of event predicates associated with a relation; each triple is assigned significance scores based on our measures. From this resource, vector-based representations of pairs of predicates can be induced and used as lexical semantic features to build models for external applications. We assess the potential of these representations for several applications. Regarding discourse analysis, the tasks of predicting attachment of discourse units, as well as predicting the specific discourse relation linking them, are investigated. Using only features from our resource, we obtain significant improvements for both tasks in comparison to several baselines, including ones using other representations of the pairs of predicates. We also propose to define optimal sets of connectives better suited for large corpus applications by performing a dimension reduction in the space of the connectives, instead of using manually composed groups of connectives corresponding to predefined relations. Another promising application pursued in this thesis concerns relations between semantic frames (e.g. FrameNet): the resource can be used to enrich this sparse structure by providing candidate relations between verbal frames, based on associations between their verbs. These diverse applications aim to demonstrate the promising contributions provided by our approach, namely allowing the unsupervised extraction of typed semantic relations.

Cette thèse, organisée en deux parties indépendantes, a pour objet la sémantique distributionnelle et la sélection de variables. Dans la première partie, nous introduisons une nouvelle méthode pour l'apprentissage de représentations de mots à partir de grandes quantités de texte brut. Cette méthode repose sur un modèle probabiliste de la phrase, utilisant modèle de Markov caché et arbre de dépendance. Nous présentons un algorithme efficace pour réaliser l'inférence et l'apprentissage dans un tel modèle, fondé sur l'algorithme EM en ligne et la propagation de message approchée. Nous évaluons les modèles obtenus sur des taches intrinsèques, telles que prédire des jugements de similarité humains ou catégoriser des mots et deux taches extrinsèques~: la reconnaissance d'entités nommées et l'étiquetage en supersens. Dans la seconde partie, nous introduisons, dans le contexte des modèles linéaires, une nouvelle pénalité pour la sélection de variables en présence de prédicteurs fortement corrélés. Cette pénalité, appelée trace Lasso, utilise la norm trace des prédicteurs sélectionnés, qui est une relaxation convexe de leur rang, comme critère de complexité. Le trace Lasso interpole les normes $\ell_1$ et $\ell_2$. En particulier, lorsque tous les prédicteurs sont orthogonaux, il est égal à la norme $\ell_1$, tandis que lorsque tous les prédicteurs sont égaux, il est égal à la norme $\ell_2$. Nous proposons deux algorithmes pour calculer la solution du problème de régression aux moindres carrés regularisé par le trace Lasso et réalisons des expériences sur des données synthétiques. / This thesis, which is organized in two independent parts, presents work on distributional semantics and on variable selection. In the first part, we introduce a new method for learning good word representations using large quantities of unlabeled sentences. The method is based on a probabilistic model of sentence, using a hidden Markov model and a syntactic dependency tree. The latent variables, which correspond to the nodes of the dependency tree, aim at capturing the meanings of the words. We develop an efficient algorithm to perform inference and learning in those models, based on online EM and approximate message passing. We then evaluate our models on intrinsic tasks such as predicting human similarity judgements or word categorization, and on two extrinsic tasks: named entity recognition and supersense tagging. In the second part, we introduce, in the context of linear models, a new penalty function to perform variable selection in the case of highly correlated predictors. This penalty, called the trace Lasso, uses the trace norm of the selected predictors, which is a convex surrogate of their rank, as the criterion of model complexity. The trace Lasso interpolates between the $\ell_1$-norm and $\ell_2$-norm. In particular, it is equal to the $\ell_1$-norm if all predictors are orthogonal and to the $\ell_2$-norm if all predictors are equal. We propose two algorithms to compute the solution of least-squares regression regularized by the trace Lasso, and perform experiments on synthetic datasets to illustrate the behavior of the trace Lasso.

A central operation of users of the text analysis tool Gavagai Explorer is to look through a list of words and arrange them in groups. This thesis explores the use of word clustering to automatically arrange the words in groups intended to help users. A new word clustering algorithm is introduced, which attempts to produce word clusters tailored to be small enough for a user to quickly grasp the common theme of the words. The proposed algorithm computes similarities among words using word embeddings, and clusters them using hierarchical graph clustering. Multiple variants of the algorithm are evaluated in an unsupervised manner by analysing the clusters they produce when applied to 110 data sets previously analysed by users of Gavagai Explorer. A supervised evaluation is performed to compare clusters to the groups of words previously created by users of Gavagai Explorer. Results show that it was possible to choose a set of hyperparameters deemed to perform well across most data sets in the unsupervised evaluation. These hyperparameters also performed among the best on the supervised evaluation. It was concluded that the choice of word embedding and graph clustering algorithm had little impact on the behaviour of the algorithm. Rather, limiting the maximum size of clusters and filtering out similarities between words had a much larger impact on behaviour.