abstract: In recent years, several methods have been proposed to encode sentences into fixed length continuous vectors called sentence representation or sentence embedding. With the recent advancements in various deep learning methods applied in Natural Language Processing (NLP), these representations play a crucial role in tasks such as named entity recognition, question answering and sentence classification. Traditionally, sentence vector representations are learnt from its constituent word representations, also known as word embeddings. Various methods to learn the distributed representation (embedding) of words have been proposed using the notion of Distributional Semantics, i.e. “meaning of a word is characterized by the company it keeps”. However, principle of compositionality states that meaning of a sentence is a function of the meanings of words and also the way they are syntactically combined. In various recent methods for sentence representation, the syntactic information like dependency or relation between words have been largely ignored. In this work, I have explored the effectiveness of sentence representations that are composed of the representation of both, its constituent words and the relations between the words in a sentence. The word and relation embeddings are learned based on their context. These general-purpose embeddings can also be used as off-the- shelf semantic and syntactic features for various NLP tasks. Similarity Evaluation tasks was performed on two datasets showing the usefulness of the learned word embeddings. Experiments were conducted on three different sentence classification tasks showing that our sentence representations outperform the original word-based sentence representations, when used with the state-of-the-art Neural Network architectures. / Dissertation/Thesis / Masters Thesis Computer Science 2017

Nowadays, knowledge bases are used more and more in Semantic Web tasks, such as knowledge acquisition (Hellmann et al., 2013), disambiguation (Garcia et al., 2009) and named entity corpus construction (Hahm et al., 2014), to name a few. DBpedia is playing a central role on the linked open data cloud; therefore, the quality of this knowledge base is becoming a central point of focus. However, there are some issues with the quality of DBpedia. In particular, DBpedia suffers from three major types of problems: a) invalid types for entities, b) missing types for entities, and c) invalid entities in the resources’ description. In order to enhance the quality of DBpedia, it is important to detect these invalid types and resources, as well as complete missing types. The three main goals of this thesis are: a) invalid entity type detection in order to solve the problem of invalid DBpedia types for entities, b) automatic detection of the types of entities in order to solve the problem of missing DBpedia types for entities, and c) invalid entity detection in order to solve the problem of invalid entities in the resource description of a DBpedia entity. We compare several methods for the detection of invalid types, automatic typing of entities, and invalid entities detection in the resource descriptions. In particular, we compare different classification and clustering algorithms based on various sets of features: entity embedding features (Skip-gram and CBOW models) and traditional n-gram features. We present evaluation results for 358 DBpedia classes extracted from the DBpedia ontology. The main contribution of this work consists of the development of automatic invalid type detection, automatic entity typing, and automatic invalid entity detection methods using clustering and classification. Our results show that entity embedding models usually perform better than n-gram models, especially the Skip-gram embedding model.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第22806号 / 情博第736号 / 新制||情||126(附属図書館) / 京都大学大学院情報学研究科社会情報学専攻 / (主査)教授 吉川 正俊, 教授 森 信介, 教授 田島 敬史 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

With the increasing amount of digital data in the form of unstructured text, the importance of natural language processing (NLP) increases. The most suc- cessful technologies of recent years are deep neural networks. This work applies the state-of-the-art methods, namely transfer learning of Bidirectional Encoders Representations from Transformers (BERT), on three Czech NLP tasks: part- of-speech tagging, lemmatization and sentiment analysis. We applied BERT model with a simple classification head on three Czech sentiment datasets: mall, facebook, and csfd, and we achieved state-of-the-art results. We also explored several possible architectures for tagging and lemmatization and obtained new state-of-the-art results in both tagging and lemmatization with fine-tunning ap- proach on data from Prague Dependency Treebank. Specifically, we achieved accuracy 98.57% for tagging, 99.00% for lemmatization, and 98.19% for joint accuracy of both tasks. Best models for all tasks are publicly available. 1

There has been an explosion of network data in the physical, chemical, biological, computational, and social sciences in the last few decades. Node embeddings, i.e., Euclidean-space representations of nodes in a network, make it possible to apply to network data, tools and algorithms from multivariate statistics and machine learning that were developed for Euclidean-space data. Random walk node embeddings are a class of recently developed node embedding techniques where the vector representations are learned by optimizing objective functions involving skip-bigram statistics computed from random walks on the network. They have been applied to many supervised learning problems such as link prediction and node classification and have demonstrated state-of-the-art performance. Yet, their properties remain poorly understood. This dissertation studies random walk based node embeddings in an unsupervised setting within the context of capturing hidden block structure in the network, i.e., learning node representations that reflect their patterns of adjacencies to other nodes. This doctoral research (i) Develops VEC, a random walk based unsupervised node embedding algorithm, and a series of relaxations, and experimentally validates their performance for the community detection problem under the Stochastic Block Model (SBM). (ii) Characterizes the ergodic limits of the embedding objectives to create non-randomized versions. (iii) Analyzes the embeddings for expected SBM networks and establishes certain concentration properties of the limiting ergodic objective in the large network asymptotic regime. Comprehensive experimental results on real world and SBM random networks are presented to illustrate and compare the distributional and block-structure properties of node embeddings generated by VEC and related algorithms. As a step towards theoretical understanding, it is proved that for the variants of VEC with ergodic limits and convex relaxations, the embedding Grammian of the expected network of a two-community SBM has rank at most 2. Further experiments reveal that these extensions yield embeddings whose distribution is Gaussian-like, centered at the node embeddings of the expected network within each community, and concentrate in the linear degree-scaling regime as the number of nodes increases. / 2023-09-24T00:00:00Z

Du point de vue du traitement automatique des langues (TAL), l’extraction des événements dans les textes est la forme la plus complexe des processus d’extraction d’information, qui recouvrent de façon plus générale l’extraction des entités nommées et des relations qui les lient dans les textes. Le cas des événements est particulièrement ardu car un événement peut être assimilé à une relation n-aire ou à une configuration de relations. Alors que la recherche en extraction d’information a largement bénéficié des jeux de données étiquetés manuellement pour apprendre des modèles permettant l’analyse des textes, la disponibilité de ces ressources reste un problème important. En outre, de nombreuses approches en extraction d’information fondées sur l’apprentissage automatique reposent sur la possibilité d’extraire à partir des textes de larges en sembles de traits définis manuellement grâce à des outils de TAL élaborés. De ce fait, l’adaptation à un nouveau domaine constitue un défi supplémentaire. Cette thèse présente plusieurs stratégies pour améliorer la performance d’un système d’extraction d’événements en utilisant des approches fondées sur les réseaux de neurones et en exploitant les propriétés morphologiques, syntaxiques et sémantiques des plongements de mots. Ceux-ci ont en effet l’avantage de ne pas nécessiter une modélisation a priori des connaissances du domaine et de générer automatiquement un ensemble de traits beaucoup plus vaste pour apprendre un modèle. Nous avons proposé plus spécifiquement différents modèles d’apprentissage profond pour les deux sous-tâches liées à l’extraction d’événements : la détection d’événements et la détection d’arguments. La détection d’événements est considérée comme une sous-tâche importante de l’extraction d’événements dans la mesure où la détection d’arguments est très directement dépendante de son résultat. La détection d’événements consiste plus précisément à identifier des instances d’événements dans les textes et à les classer en types d’événements précis. En préalable à l’introduction de nos nouveaux modèles, nous commençons par présenter en détail le modèle de l’état de l’art qui en constitue la base. Des expériences approfondies sont menées sur l’utilisation de différents types de plongements de mots et sur l’influence des différents hyperparamètres du modèle en nous appuyant sur le cadre d’évaluation ACE 2005, standard d’évaluation pour cette tâche. Nous proposons ensuite deux nouveaux modèles permettant d’améliorer un système de détection d’événements. L’un permet d’augmenter le contexte pris en compte lors de la prédiction d’une instance d’événement (déclencheur d’événement) en utilisant un contexte phrastique, tandis que l’autre exploite la structure interne des mots en profitant de connaissances morphologiques en apparence moins nécessaires mais dans les faits importantes. Nous proposons enfin de reconsidérer la détection des arguments comme une extraction de relation d’ordre supérieur et nous analysons la dépendance de cette détection vis-à-vis de la détection d’événements. / With the increasing amount of data and the exploding number data sources, the extraction of information about events, whether from the perspective of acquiring knowledge or from a more directly operational perspective, becomes a more and more obvious need. This extraction nevertheless comes up against a recurring difficulty: most of the information is present in documents in a textual form, thus unstructured and difficult to be grasped by the machine. From the point of view of Natural Language Processing (NLP), the extraction of events from texts is the most complex form of Information Extraction (IE) techniques, which more generally encompasses the extraction of named entities and relationships that bind them in the texts. The event extraction task can be represented as a complex combination of relations linked to a set of empirical observations from texts. Compared to relations involving only two entities, there is, therefore, a new dimension that often requires going beyond the scope of the sentence, which constitutes an additional difficulty. In practice, an event is described by a trigger and a set of participants in that event whose values are text excerpts. While IE research has benefited significantly from manually annotated datasets to learn patterns for text analysis, the availability of these resources remains a significant problem. These datasets are often obtained through the sustained efforts of research communities, potentially complemented by crowdsourcing. In addition, many machine learning-based IE approaches rely on the ability to extract large sets of manually defined features from text using sophisticated NLP tools. As a result, adaptation to a new domain is an additional challenge. This thesis presents several strategies for improving the performance of an Event Extraction (EE) system using neural-based approaches exploiting morphological, syntactic, and semantic properties of word embeddings. These have the advantage of not requiring a priori modeling domain knowledge and automatically generate a much larger set of features to learn a model. More specifically, we proposed different deep learning models for two sub-tasks related to EE: event detection and argument detection and classification. Event Detection (ED) is considered an important subtask of event extraction since the detection of arguments is very directly dependent on its outcome. ED specifically involves identifying instances of events in texts and classifying them into specific event types. Classically, the same event may appear as different expressions and these expressions may themselves represent different events in different contexts, hence the difficulty of the task. The detection of the arguments is based on the detection of the expression considered as triggering the event and ensures the recognition of the participants of the event. Among the difficulties to take into account, it should be noted that an argument can be common to several events and that it does not necessarily identify with an easily recognizable named entity. As a preliminary to the introduction of our proposed models, we begin by presenting in detail a state-of-the-art model which constitutes the baseline. In-depth experiments are conducted on the use of different types of word embeddings and the influence of the different hyperparameters of the model using the ACE 2005 evaluation framework, a standard evaluation for this task. We then propose two new models to improve an event detection system. One allows increasing the context taken into account when predicting an event instance by using a sentential context, while the other exploits the internal structure of words by taking advantage of seemingly less obvious but essentially important morphological knowledge. We also reconsider the detection of arguments as a high-order relation extraction and we analyze the dependence of arguments on the ED task.

Word embeddings are utilized in various natural language processing tasks. Although effective in helping computers learn linguistic patterns employed in natural language, word embeddings also tend to learn unwanted word associations. This affects the performance of NLP tasks, as unwanted word associations propagate and amplify biases. Current word association evaluation methods for word embeddings do not account for changes in word embedding models and training corpora, when creating the rubric for word association evaluation. Current literature also lacks a consistent training and evaluation protocol for comparison of word associations across varying word embedding models and varying training corpora. In order to address this gap in prior literature, this research aims to evaluate different types of word associations, not limited to gender, racial or religious attributes, incorporating and evaluating the diachronic and variable nature of words over text data collected over a period of 200 years. This thesis introduces a framework to track changes in word associations between neutral words (proper nouns) and attributes (adjectives), across different word embedding models, over a temporal dimension, by evaluating clustering tendencies between neutral words (proper nouns) and attributive words (adjectives) over five different word embedding frameworks: Word2vec (CBOW), Word2vec (Skip-gram), GloVe, fastText (CBOW) and fastText (Skip-gram) and 20 decades of text data from 1810s to 2000s. <a>Finally, various cluster level and corpus level measurements will be compared across aforementioned word embedding frameworks, to find how</a> word associations evolve with changes in the embedding model and the training corpus.

Word embeddings are useful in many tasks in Natural Language Processing and Information Retrieval, such as text mining and classification, sentiment analysis, sentence completion, or dictionary construction. Word2vec and its predecessor fastText, both well-known models to produce word embeddings, are powerful techniques to study the syntactic and semantic relations between words by representing them in a low-dimensional vector. By applying algebraic operations on these vectors semantic relationships such as word analogies, gender-inflections, or geographical relationships can be easily recovered. The aim of this work is to investigate how word embeddings could be utilized to augment and enrich queries in DBMSs, e.g. to compare text values according to their semantic relation or to group rows according to the similarity of their text values. For this purpose, we use pre-trained word embedding models of large text corpora such as Wikipedia. By exploiting this external knowledge during query processing we are able to apply inductive reasoning on text values. Thereby, we reduce the demand for explicit knowledge in database systems. In the context of the IMDB database schema, this allows for example to query movies that are semantically close to genres such as historical fiction or road movie without maintaining this information. Another example query is sketched in Listing 1, that returns the top-3 nearest neighbors (NN) of each movie in IMDB. Given the movie “Godfather” as input this results in “Scarface”, “Goodfellas” and “Untouchables”.

This dissertation advances the state-of-the-art in information retrieval (IR) based automatic bug localization for large software systems. We present techniques from three generations of IR based bug localization and compare their performances on our large and diverse bug localization dataset --- the Bugzbook dataset. The three generations span over fifteen years of research in mining software repositories for bug localization and include: (1) the generation of simple bag-of-words (BoW) based techniques, (2) the generation in which software-centric information such as bug and code change histories as well as structured information embedded in bug reports and code files are exploited to improve retrieval, and (3) the third and most recent generation in which order and semantic relationships between terms are modeled to improve the performance of bug localization systems. The dissertation also presents a novel technique called SCOR (Source Code Retrieval with Semantics and Order) which combines Markov Random Fields (MRF) based term-term ordering dependencies with semantic word vectors obtained from neural network based word embedding algorithms, such as word2vec, to better localize bugs in code files. The results presented in this dissertation show that while term-term ordering and semantic relationships significantly improve the performance when they are modeled separately in retrieval systems, the best precisions in retrieval are obtained when they are modeled together in a single retrieval system. We also show that the semantic representations of software terms learned by training the word embedding algorithm on a corpus of software repositories can be used to perform search in new software code repositories not present in the training corpus of the word embedding algorithm.<br>

Predictive Maintenance (PM) has been increasingly adopted in the Automotive industry, in the recent decades along with conventional approaches such as the Preventive Maintenance and Diagnostic/Corrective Maintenance, since it provides many advantages to estimate the failure before the actual occurrence proactively, and also being adaptive to the present status of the vehicle, in turn allowing flexible maintenance schedules for efficient repair or replacing of faulty components. PM necessitates the storage and analysis of large amounts of sensor data. This requirement can be a challenge in deploying this method on-board the vehicles due to the limited storage and computational power on the hardware of the vehicle. Hence, this thesis seeks to obtain low dimensional descriptive features from high dimensional data using Representation Learning. This low dimensional representation will be used for predicting vehicle faults, specifically Turbocharger related failures. Since the Logged Vehicle Data (LVD) was base on all the data utilized in this thesis, it allowed for the evaluation of large populations of trucks without requiring additional measuring devices and facilities. The gradual degradation methodology is considered for describing vehicle condition, which allows for modeling the malfunction/ failure as a continuous process rather than a discrete flip from healthy to an unhealthy state. This approach eliminates the challenge of data imbalance of healthy and unhealthy samples. Two important hypotheses are presented. Firstly, Parallel StackedClassical Autoencoders would produce better representations com-pared to individual Autoencoders. Secondly, employing Learned Em-beddings on Categorical Variables would improve the performance of the Dimensionality reduction. Based on these hypotheses, a model architecture is proposed and is developed on the LVD. The model is shown to achieve good performance, and in close standards to the previous state-of-the-art research. This thesis, finally, illustrates the potential to apply parallel stacked architectures with Learned Embeddings for the Categorical features, and a combination of feature selection and extraction for numerical features, to predict the Remaining Useful Life (RUL) of a vehicle, in the context of the Turbocharger. A performance improvement of 21.68% with respect to the Mean Absolute Error (MAE) loss with an 80.42% reduction in the size of data was observed.