Tesis sobre el tema "Cross lingual text classification"

Automated classification of texts into genres can benefit NLP applications, in that the structure, location and even interpretation of information within a text are dictated by its genre. Cross-lingual methods promise such benefits to languages which lack genre-annotated training data. While there has been work on genre classification for over two decades, none has considered cross-lingual methods before the start of this project. My research aims to fill this gap. It follows previous approaches to monolingual genre classification that exploit simple, low-level text features, many of which can be extracted in different languages and have similar functions. This contrasts with work on cross-lingual topic or sentiment classification of texts that typically use word frequencies as features. These have been shown to have limited use when it comes to genres. Many such methods also assume cross-lingual resources, such as machine translation, which limits the range of their application. A selection of these approaches are used as baselines in my experiments. I report the results of two semi-supervised methods for exploiting genre-labelled source language texts and unlabelled target language texts. The first is a relatively simple algorithm that bridges the language gap by exploiting cross-lingual features and then iteratively re-trains a classification model on previously predicted target texts. My results show that this approach works well where only few cross-lingual resources are available and texts are to be classified into broad genre categories. It is also shown that further improvements can be achieved through multi-lingual training or cross-lingual feature selection if genre-annotated texts are available in several source languages. The second is a variant of the label propagation algorithm. This graph-based classifier learns genre-specific feature set weights from both source and target language texts and uses them to adjust the propagation channels for each text. This allows further feature sets to be added as additional resources, such as Part of Speech taggers, become available. While the method performs well even with basic text features, it is shown to benefit from additional feature sets. Results also indicate that it handles fine-grained genre classes better than the iterative re-labelling method.

Cross-lingual text classification is an important task due to the globalization and the increased availability of multilingual data. This thesis explores the method of implementing cross-lingual classification on Swedish and English medical corpora. Specifically, this the- sis explores the simple convolutional neural network (CNN) with MUSE pre-trained word embeddings to approach binary classification of sublanguages (“lay” and “specialized”) from Swedish healthcare texts to English healthcare texts. MUSE is a library that provides state-of-the-art multilingual word embeddings and large-scale high-quality bilingual dictionaries. The thesis presents experiments with imbalanced and balanced class distribution on training data and test data to examine the effect of class distribution, and also examine the influences of clean test dataset and noisy test dataset. The results show that balanced distribution of classes in training data performs significantly better than the training data with imbalanced class distribution, and clean test data gives the benefit of transferring the labels from one language to another. The thesis also compares the performance of the simple convolutional neural network model with the Naive Bayes baseline. Results show that on this task a simple Naive Bayes classifier based on bag-of-words translated using MUSE English-Swedish dictionary outperforms a simple CNN model based on MUSE pre-trained word embeddings in several experimental settings.

Emotions can be defined as a natural, instinctive state of mind arising from one’s circumstances, mood, and relationships with others. It has long been a question to be answered by psychology that how and what is it that humans feel. Enabling computers to recognize human emotions has been an of interest to researchers since 1990s (Picard et al., 1995). Ever since, this area of research has grown significantly and emotion detection is becoming an important component in many natural language processing tasks. Several theories exist for defining emotions and are chosen by researchers according to their needs. For instance, according to appraisal theory, a psychology theory, emotions are produced by our evaluations (appraisals or estimates) of events that cause a specific reaction in different people. Some emotions are easy and universal, while others are complex and nuanced. Emotion classification is generally the process of labeling a piece of text with one or more corresponding emotion labels. Psychologists have developed numerous models and taxonomies of emotions. The model or taxonomy depends on the problem, and thorough study is often required to select the best model. Early studies of emotion classification focused on building computational models to classify basic emotion categories. In recent years, increasing volumes of social media and the digitization of data have opened a new horizon in this area of study, where emotion classification is a key component of applications, including mood and behavioral studies, as well as disaster relief, amongst many other applications. Sophisticated models have been built to detect and classify emotion in text, but few analyze how well a model is able to learn emotion cues. The ability to learn emotion cues properly and be able to generalize this learning is very important. This work investigates the robustness of emotion classification approaches across genres and languages, with a focus on quantifying how well state-of-the-art models are able to learn emotion cues. First, we use multi-task learning and hierarchical models to build emotion models that were trained on data combined from multiple genres. Our hypothesis is that a multi-genre, noisy training environment will help the classifier learn emotion cues that are prevalent across genres. Second, we explore splitting text (i.e. sentence) into its clauses and testing whether the model’s performance improves. Emotion analysis needs fine-grained annotation and clause-level annotation can be beneficial to design features to improve emotion detection performance. Intuitively, clause-level annotations may help the model focus on emotion cues, while ignoring irrelevant portions of the text. Third, we adopted a transfer learning approach for cross-lingual/genre emotion classification to focus the classifier’s attention on emotion cues which are consistent across languages. Fourth, we empirically show how to combine different genres to be able to build robust models that can be used as source models for emotion transfer to low-resource target languages. Finally, this study involved curating and re-annotating popular emotional data sets in different genres, and annotating a multi-genre corpus of Persian tweets and news, and generating a collection of emotional sentences for a low-resource language, Azerbaijani, a language spoken in the north west of Iran.

The written word plays an important role in the reinforcement of gender stereotypes, especially in texts of a more formal character. Wikipedia biographies have a lot of information about famous people, but do they describe men and women with different kinds of words? This thesis aims to evaluate and explore a method for gender classification of text. In this study, two machine learning classifiers, Random Forest (RF) and Support Vector Machine (SVM), are applied to the gender classification of Wikipedia biographies in two languages, English and French. Their performance is evaluated and compared. The 500 most important words (features) are listed for each of the classifiers.A short review is given on the theoretic foundations of text classification, and a detailed description on how the datasets are built, what tools are used, and why. The datasets used are built from the first 5 paragraphs in each biography, with only nouns, verbs, adjectives and adverbs remaining. Feature ranking is also applied, where the top tenth of the features are kept.Performance is measured using the F0:5-score. The comparison shows that the RF and SVM classifiers' performance are close to each other, but that the classifiers perform worse on the French set than on the English. Initial performance scores range from 0.82 to 0.86, but they drop drastically when the most important features are removed from the set. A majority of the top most important features are nouns related to career and family roles, in both languages.The results show that there are indeed some semantic differences in language depending on the gender of the person described. Whether these depend on the writers' biased views, an unequal gender distribution of real world contexts, such as careers, or if these differences depend on how the datasets were built, is not clear.

Master of Science
Department of Computing and Information Sciences
Doina Caragea
Social media affects our daily lives. It is one of the first sources for finding breaking news. In particular, Twitter is one of the popular social media platforms, with around 330 million monthly users. From local events such as Fake Patty's Day to across the world happenings - Twitter gets there first. During a disaster, tweets can be used to post warnings, status of available medical and food supply, emergency personnel, and updates. Users were practically tweeting about the Hurricane Sandy, despite lack of network during the storm. Analysis of these tweets can help monitor the disaster, plan and manage the crisis, and aid in research. In this research, we use the publicly available tweets posted during several disasters and identify the relevant tweets. As the languages in the datasets are different, Bing translation API has been used to detect and translate the tweets. The translations are then, used as training datasets for supervised machine learning algorithms. Supervised learning is the process of learning from a labeled training dataset. This learned classifier can then be used to predict the correct output for any valid input. When trained to more observations, the algorithm improves its predictive performance.

With the growing amount of data generated in large heterogeneous repositories (such as the Word Wide Web, corporate repositories, citation databases), there is an increased need for the end users to locate relevant information efficiently. Text Classification (TC) techniques provide automated means for classifying fragments of text (phrases, paragraphs or documents) into predefined semantic types, allowing an efficient way for organising and analysing such large document collections. Current approaches to TC rely on supervised learning, which perform well on the domains on which the TC system is built, but tend to adapt poorly to different domains. This thesis presents a body of work for exploring adaptive TC techniques across hetero- geneous corpora in large repositories with the goal of finding novel ways of bridging the gap across domains. The proposed approaches rely on the exploitation of domain knowledge for the derivation of stable cross-domain features. This thesis also investigates novel ways of estimating the performance of a TC classifier, by means of domain similarity measures. For this purpose, two novel knowledge-based similarity measures are proposed that capture the usefulness of the selected cross-domain features for cross-domain TC. The evaluation of these approaches and measures is presented on real world datasets against various strong baseline methods and content-based measures used in transfer learning. This thesis explores how domain knowledge can be used to enhance the representation of documents to address the lexical gap across the domains. Given that the effectiveness of a text classifier largely depends on the availability of annotated data, this thesis explores techniques which can leverage data from social knowledge sources (such as DBpedia and Freebase). Techniques are further presented, which explore the feasibility of exploiting different semantic graph structures from knowledge sources in order to create novel cross- domain features and domain similarity metrics. The methodologies presented provide a novel representation of documents, and exploit four wide coverage knowledge sources: DBpedia, Freebase, SNOMED-CT and MeSH. The contribution of this thesis demonstrates the feasibility of exploiting domain knowl- edge for adaptive TC and domain similarity, providing an enhanced representation of docu- ments with semantic information about entities, that can indeed reduce the lexical differences between domains.

The Web is a vast repository of data, and information on almost any subject can be found with the aid of search engines. Although the Web is international, the majority of research on finding of information has a focus on languages such as English and Chinese. In this thesis, we investigate information retrieval techniques for Indonesian. Although Indonesia is the fourth most populous country in the world, little attention has been given to search of Indonesian documents. Stemming is the process of reducing morphological variants of a word to a common stem form. Previous research has shown that stemming is language-dependent. Although several stemming algorithms have been proposed for Indonesian, there is no consensus on which gives better performance. We empirically explore these algorithms, showing that even the best algorithm still has scope for improvement. We propose novel extensions to this algorithm and develop a new Indonesian stemmer, and show that these can improve stemming correctness by up to three percentage points; our approach makes less than one error in thirty-eight words. We propose a range of techniques to enhance the performance of Indonesian information retrieval. These techniques include: stopping; sub-word tokenisation; and identification of proper nouns; and modifications to existing similarity functions. Our experiments show that many of these techniques can increase retrieval performance, with the highest increase achieved when we use grams of size five to tokenise words. We also present an effective method for identifying the language of a document; this allows various information retrieval techniques to be applied selectively depending on the language of target documents. We also address the problem of automatic creation of parallel corpora --- collections of documents that are the direct translations of each other --- which are essential for cross-lingual information retrieval tasks. Well-curated parallel corpora are rare, and for many languages, such as Indonesian, do not exist at all. We describe algorithms that we have developed to automatically identify parallel documents for Indonesian and English. Unlike most current approaches, which consider only the context and structure of the documents, our approach is based on the document content itself. Our algorithms do not make any prior assumptions about the documents, and are based on the Needleman-Wunsch algorithm for global alignment of protein sequences. Our approach works well in identifying Indonesian-English parallel documents, especially when no translation is performed. It can increase the separation value, a measure to discriminate good matches of parallel documents from bad matches, by approximately ten percentage points. We also investigate the applicability of our identification algorithms for other languages that use the Latin alphabet. Our experiments show that, with minor modifications, our alignment methods are effective for English-French, English-German, and French-German corpora, especially when the documents are not translated. Our technique can increase the separation value for the European corpus by up to twenty-eight percentage points. Together, these results provide a substantial advance in understanding techniques that can be applied for effective Indonesian text retrieval.

Une des promesses des plateformes de réseaux sociaux (comme Twitter et Facebook) est de fournir un endroit sûr pour que les utilisateurs puissent partager leurs opinions et des informations. Cependant, l’augmentation des comportements abusifs, comme le harcèlement en ligne ou la présence de discours de haine, est bien réelle. Dans cette thèse, nous nous concentrons sur le discours de haine, l'un des phénomènes les plus préoccupants concernant les réseaux sociaux.Compte tenu de sa forte progression et de ses graves effets négatifs, les institutions, les plateformes de réseaux sociaux et les chercheurs ont tenté de réagir le plus rapidement possible. Les progrès récents des algorithmes de traitement automatique du langage naturel (NLP) et d'apprentissage automatique (ML) peuvent être adaptés pour développer des méthodes automatiques de détection des discours de haine dans ce domaine.Le but de cette thèse est d'étudier le problème du discours de haine et de la détection des propos injurieux dans les réseaux sociaux. Nous proposons différentes approches dans lesquelles nous adaptons des modèles avancés d'apprentissage par transfert (TL) et des techniques de NLP pour détecter automatiquement les discours de haine et les contenus injurieux, de manière monolingue et multilingue.La première contribution concerne uniquement la langue anglaise. Tout d'abord, nous analysons le contenu textuel généré par les utilisateurs en introduisant un nouveau cadre capable de catégoriser le contenu en termes de similarité basée sur différentes caractéristiques. En outre, en utilisant l'API Perspective de Google, nous mesurons et analysons la « toxicité » du contenu. Ensuite, nous proposons une approche TL pour l'identification des discours de haine en utilisant une combinaison du modèle non supervisé pré-entraîné BERT (Bidirectional Encoder Representations from Transformers) et de nouvelles stratégies supervisées de réglage fin. Enfin, nous étudions l'effet du biais involontaire dans notre modèle pré-entraîné BERT et proposons un nouveau mécanisme de généralisation dans les données d'entraînement en repondérant les échantillons puis en changeant les stratégies de réglage fin en termes de fonction de perte pour atténuer le biais racial propagé par le modèle. Pour évaluer les modèles proposés, nous utilisons deux datasets publics provenant de Twitter.Dans la deuxième contribution, nous considérons un cadre multilingue où nous nous concentrons sur les langues à faibles ressources dans lesquelles il n'y a pas ou peu de données annotées disponibles. Tout d'abord, nous présentons le premier corpus de langage injurieux en persan, composé de 6 000 messages de micro-blogs provenant de Twitter, afin d'étudier la détection du langage injurieux. Après avoir annoté le corpus, nous réalisons étudions les performances des modèles de langages pré-entraînés monolingues et multilingues basés sur des transformeurs (par exemple, ParsBERT, mBERT, XLM-R) dans la tâche en aval. De plus, nous proposons un modèle d'ensemble pour améliorer la performance de notre modèle. Enfin, nous étendons notre étude à un problème d'apprentissage multilingue de type " few-shot ", où nous disposons de quelques données annotées dans la langue cible, et nous adaptons une approche basée sur le méta-apprentissage pour traiter l'identification des discours de haine et du langage injurieux dans les langues à faibles ressources
The great promise of social media platforms (e.g., Twitter and Facebook) is to provide a safe place for users to communicate their opinions and share information. However, concerns are growing that they enable abusive behaviors, e.g., threatening or harassing other users, cyberbullying, hate speech, racial and sexual discrimination, as well. In this thesis, we focus on hate speech as one of the most concerning phenomenon in online social media.Given the high progression of online hate speech and its severe negative effects, institutions, social media platforms, and researchers have been trying to react as quickly as possible. The recent advancements in Natural Language Processing (NLP) and Machine Learning (ML) algorithms can be adapted to develop automatic methods for hate speech detection in this area.The aim of this thesis is to investigate the problem of hate speech and offensive language detection in social media, where we define hate speech as any communication criticizing a person or a group based on some characteristics, e.g., gender, sexual orientation, nationality, religion, race. We propose different approaches in which we adapt advanced Transfer Learning (TL) models and NLP techniques to detect hate speech and offensive content automatically, in a monolingual and multilingual fashion.In the first contribution, we only focus on English language. Firstly, we analyze user-generated textual content to gain a brief insight into the type of content by introducing a new framework being able to categorize contents in terms of topical similarity based on different features. Furthermore, using the Perspective API from Google, we measure and analyze the toxicity of the content. Secondly, we propose a TL approach for identification of hate speech by employing a combination of the unsupervised pre-trained model BERT (Bidirectional Encoder Representations from Transformers) and new supervised fine-tuning strategies. Finally, we investigate the effect of unintended bias in our pre-trained BERT based model and propose a new generalization mechanism in training data by reweighting samples and then changing the fine-tuning strategies in terms of the loss function to mitigate the racial bias propagated through the model. To evaluate the proposed models, we use two publicly available datasets from Twitter.In the second contribution, we consider a multilingual setting where we focus on low-resource languages in which there is no or few labeled data available. First, we present the first corpus of Persian offensive language consisting of 6k micro blog posts from Twitter to deal with offensive language detection in Persian as a low-resource language in this domain. After annotating the corpus, we perform extensive experiments to investigate the performance of transformer-based monolingual and multilingual pre-trained language models (e.g., ParsBERT, mBERT, XLM-R) in the downstream task. Furthermore, we propose an ensemble model to boost the performance of our model. Then, we expand our study into a cross-lingual few-shot learning problem, where we have a few labeled data in target language, and adapt a meta-learning based approach to address identification of hate speech and offensive language in low-resource languages

Natural Language Processing (NLP) is a field of computer science, artificial intelligence, and computational linguistics concerned with the interactions between computers and human languages. One of its most challenging aspects involves enabling computers to derive meaning from human natural language. To do so, several meaning or context representations have been proposed with competitive performance. However, these representations still have room for improvement when working in a cross-domain or cross-language scenario. In this thesis we study the use of knowledge graphs as a cross-domain and cross-language representation of text and its meaning. A knowledge graph is a graph that expands and relates the original concepts belonging to a set of words. We obtain its characteristics using a wide-coverage multilingual semantic network as knowledge base. This allows to have a language coverage of hundreds of languages and millions human-general and -specific concepts. As starting point of our research we employ knowledge graph-based features - along with other traditional ones and meta-learning - for the NLP task of single- and cross-domain polarity classification. The analysis and conclusions of that work provide evidence that knowledge graphs capture meaning in a domain-independent way. The next part of our research takes advantage of the multilingual semantic network and focuses on cross-language Information Retrieval (IR) tasks. First, we propose a fully knowledge graph-based model of similarity analysis for cross-language plagiarism detection. Next, we improve that model to cover out-of-vocabulary words and verbal tenses and apply it to cross-language document retrieval, categorisation, and plagiarism detection. Finally, we study the use of knowledge graphs for the NLP tasks of community questions answering, native language identification, and language variety identification. The contributions of this thesis manifest the potential of knowledge graphs as a cross-domain and cross-language representation of text and its meaning for NLP and IR tasks. These contributions have been published in several international conferences and journals.
El Procesamiento del Lenguaje Natural (PLN) es un campo de la informática, la inteligencia artificial y la lingüística computacional centrado en las interacciones entre las máquinas y el lenguaje de los humanos. Uno de sus mayores desafíos implica capacitar a las máquinas para inferir el significado del lenguaje natural humano. Con este propósito, diversas representaciones del significado y el contexto han sido propuestas obteniendo un rendimiento competitivo. Sin embargo, estas representaciones todavía tienen un margen de mejora en escenarios transdominios y translingües. En esta tesis estudiamos el uso de grafos de conocimiento como una representación transdominio y translingüe del texto y su significado. Un grafo de conocimiento es un grafo que expande y relaciona los conceptos originales pertenecientes a un conjunto de palabras. Sus propiedades se consiguen gracias al uso como base de conocimiento de una red semántica multilingüe de amplia cobertura. Esto permite tener una cobertura de cientos de lenguajes y millones de conceptos generales y específicos del ser humano. Como punto de partida de nuestra investigación empleamos características basadas en grafos de conocimiento - junto con otras tradicionales y meta-aprendizaje - para la tarea de PLN de clasificación de la polaridad mono- y transdominio. El análisis y conclusiones de ese trabajo muestra evidencias de que los grafos de conocimiento capturan el significado de una forma independiente del dominio. La siguiente parte de nuestra investigación aprovecha la capacidad de la red semántica multilingüe y se centra en tareas de Recuperación de Información (RI). Primero proponemos un modelo de análisis de similitud completamente basado en grafos de conocimiento para detección de plagio translingüe. A continuación, mejoramos ese modelo para cubrir palabras fuera de vocabulario y tiempos verbales, y lo aplicamos a las tareas translingües de recuperación de documentos, clasificación, y detección de plagio. Por último, estudiamos el uso de grafos de conocimiento para las tareas de PLN de respuesta de preguntas en comunidades, identificación del lenguaje nativo, y identificación de la variedad del lenguaje. Las contribuciones de esta tesis ponen de manifiesto el potencial de los grafos de conocimiento como representación transdominio y translingüe del texto y su significado en tareas de PLN y RI. Estas contribuciones han sido publicadas en diversas revistas y conferencias internacionales.
El Processament del Llenguatge Natural (PLN) és un camp de la informàtica, la intel·ligència artificial i la lingüística computacional centrat en les interaccions entre les màquines i el llenguatge dels humans. Un dels seus majors reptes implica capacitar les màquines per inferir el significat del llenguatge natural humà. Amb aquest propòsit, diverses representacions del significat i el context han estat proposades obtenint un rendiment competitiu. No obstant això, aquestes representacions encara tenen un marge de millora en escenaris trans-dominis i trans-llenguatges. En aquesta tesi estudiem l'ús de grafs de coneixement com una representació trans-domini i trans-llenguatge del text i el seu significat. Un graf de coneixement és un graf que expandeix i relaciona els conceptes originals pertanyents a un conjunt de paraules. Les seves propietats s'aconsegueixen gràcies a l'ús com a base de coneixement d'una xarxa semàntica multilingüe d'àmplia cobertura. Això permet tenir una cobertura de centenars de llenguatges i milions de conceptes generals i específics de l'ésser humà. Com a punt de partida de la nostra investigació emprem característiques basades en grafs de coneixement - juntament amb altres tradicionals i meta-aprenentatge - per a la tasca de PLN de classificació de la polaritat mono- i trans-domini. L'anàlisi i conclusions d'aquest treball mostra evidències que els grafs de coneixement capturen el significat d'una forma independent del domini. La següent part de la nostra investigació aprofita la capacitat\hyphenation{ca-pa-ci-tat} de la xarxa semàntica multilingüe i se centra en tasques de recuperació d'informació (RI). Primer proposem un model d'anàlisi de similitud completament basat en grafs de coneixement per a detecció de plagi trans-llenguatge. A continuació, vam millorar aquest model per cobrir paraules fora de vocabulari i temps verbals, i ho apliquem a les tasques trans-llenguatges de recuperació de documents, classificació, i detecció de plagi. Finalment, estudiem l'ús de grafs de coneixement per a les tasques de PLN de resposta de preguntes en comunitats, identificació del llenguatge natiu, i identificació de la varietat del llenguatge. Les contribucions d'aquesta tesi posen de manifest el potencial dels grafs de coneixement com a representació trans-domini i trans-llenguatge del text i el seu significat en tasques de PLN i RI. Aquestes contribucions han estat publicades en diverses revistes i conferències internacionals.
Franco Salvador, M. (2017). A Cross-domain and Cross-language Knowledge-based Representation of Text and its Meaning [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/84285
TESIS

So far, studies seeking to recognise moral foundations in texts have been relatively successful (Araque et al., 2019; Lin et al., 2018; Mooijman et al., 2017; Rezapouret al., 2019). There are, however, two issues with these studies: Firstly, it is an extensive process to gather and annotate sufficient material for training. Secondly, models are only trained and tested within the same domain. It is yet unexplored how these models for moral foundation prediction perform when tested in other domains, but from their experience with annotation, Hoover et al. (2017) describe how moral sentiments on one topic (e.g. black lives matter) might be completely different from moral sentiments on another (e.g. presidential elections). This study attempts to explore to what extent models generalise to other domains. More specifically, we focus on training on Twitter data from non-extremist sources, and testing on data from an extremist (white nationalist) forum. We conducted two experiments. In our first experiment we test whether it is possible to do cross domain classification of moral foundations. Additionally, we compare the performance of a model using the Word2Vec embeddings used in previous studies to a model using the newer BERT embeddings. We find that although the performance drops significantly on the extremist out-domain test sets, out-domain classification is not impossible. Furthermore, we find that the BERT model generalises marginally better to the out-domain test set, than the Word2Vec model. In our second experiment we attempt to improve the generalisation to extremist test data by providing contextual knowledge. Although this does not improve the model, it does show the model’s robustness against noise. Finally we suggest an alternative approach for accounting for contextual knowledge.

La présente thèse étudie la modélisation conjointe des contenus visuels et textuels extraits à partir des documents multimédias pour résoudre les problèmes intermodaux. Ces tâches exigent la capacité de ``traduire'' l'information d'une modalité vers une autre. Un espace de représentation commun, par exemple obtenu par l'Analyse Canonique des Corrélation ou son extension kernelisée est une solution généralement adoptée. Sur cet espace, images et texte peuvent être représentés par des vecteurs de même type sur lesquels la comparaison intermodale peut se faire directement.Néanmoins, un tel espace commun souffre de plusieurs déficiences qui peuvent diminuer la performance des ces tâches. Le premier défaut concerne des informations qui sont mal représentées sur cet espace pourtant très importantes dans le contexte de la recherche intermodale. Le deuxième défaut porte sur la séparation entre les modalités sur l'espace commun, ce qui conduit à une limite de qualité de traduction entre modalités. Pour faire face au premier défaut concernant les données mal représentées, nous avons proposé un modèle qui identifie tout d'abord ces informations et puis les combine avec des données relativement bien représentées sur l'espace commun. Les évaluations sur la tâche d'illustration de texte montrent que la prise en compte de ces information fortement améliore les résultats de la recherche intermodale. La contribution majeure de la thèse se concentre sur la séparation entre les modalités sur l'espace commun pour améliorer la performance des tâches intermodales. Nous proposons deux méthodes de représentation pour les documents bi-modaux ou uni-modaux qui regroupent à la fois des informations visuelles et textuelles projetées sur l'espace commun. Pour les documents uni-modaux, nous suggérons un processus de complétion basé sur un ensemble de données auxiliaires pour trouver les informations correspondantes dans la modalité absente. Ces informations complémentaires sont ensuite utilisées pour construire une représentation bi-modale finale pour un document uni-modal. Nos approches permettent d'obtenir des résultats de l'état de l'art pour la recherche intermodale ou la classification bi-modale et intermodale
This thesis investigates the joint modeling of visual and textual content of multimedia documents to address cross-modal problems. Such tasks require the ability to match information across modalities. A common representation space, obtained by eg Kernel Canonical Correlation Analysis, on which images and text can be both represented and directly compared is a generally adopted solution.Nevertheless, such a joint space still suffers from several deficiencies that may hinder the performance of cross-modal tasks. An important contribution of this thesis is therefore to identify two major limitations of such a space. The first limitation concerns information that is poorly represented on the common space yet very significant for a retrieval task. The second limitation consists in a separation between modalities on the common space, which leads to coarse cross-modal matching. To deal with the first limitation concerning poorly-represented data, we put forward a model which first identifies such information and then finds ways to combine it with data that is relatively well-represented on the joint space. Evaluations on emph{text illustration} tasks show that by appropriately identifying and taking such information into account, the results of cross-modal retrieval can be strongly improved. The major work in this thesis aims to cope with the separation between modalities on the joint space to enhance the performance of cross-modal tasks.We propose two representation methods for bi-modal or uni-modal documents that aggregate information from both the visual and textual modalities projected on the joint space. Specifically, for uni-modal documents we suggest a completion process relying on an auxiliary dataset to find the corresponding information in the absent modality and then use such information to build a final bi-modal representation for a uni-modal document. Evaluations show that our approaches achieve state-of-the-art results on several standard and challenging datasets for cross-modal retrieval or bi-modal and cross-modal classification

Nowadays communication is switching from a centralized scenario, where communication media like newspapers, radio, TV programs produce information and people are just consumers, to a completely different decentralized scenario, where everyone is potentially an information producer through the use of social networks, blogs, forums that allow a real-time worldwide information exchange. These new instruments, as a result of their widespread diffusion, have started playing an important socio-economic role. They are the most used communication media and, as a consequence, they constitute the main source of information enterprises, political parties and other organizations can rely on. Analyzing data stored in servers all over the world is feasible by means of Text Mining techniques like Sentiment Analysis, which aims to extract opinions from huge amount of unstructured texts. This could lead to determine, for instance, the user satisfaction degree about products, services, politicians and so on. In this context, this dissertation presents new Document Sentiment Classification methods based on the mathematical theory of Markov Chains. All these approaches bank on a Markov Chain based model, which is language independent and whose killing features are simplicity and generality, which make it interesting with respect to previous sophisticated techniques. Every discussed technique has been tested in both Single-Domain and Cross-Domain Sentiment Classification areas, comparing performance with those of other two previous works. The performed analysis shows that some of the examined algorithms produce results comparable with the best methods in literature, with reference to both single-domain and cross-domain tasks, in $2$-classes (i.e. positive and negative) Document Sentiment Classification. However, there is still room for improvement, because this work also shows the way to walk in order to enhance performance, that is, a good novel feature selection process would be enough to outperform the state of the art. Furthermore, since some of the proposed approaches show promising results in $2$-classes Single-Domain Sentiment Classification, another future work will regard validating these results also in tasks with more than $2$ classes.

L’objectif de cette thèse est d’étudier les sentiments dans les documents comparables. Premièrement, nous avons recueillis des corpus comparables en anglais, français et arabe de Wikipédia et d’Euronews, et nous avons aligné ces corpus au niveau document. Nous avons en plus collecté des documents d’informations des agences de presse locales et étrangères dans les langues anglaise et arabe. Les documents en anglais ont été recueillis du site de la BBC, ceux en arabe du site d’Al-Jazzera. Deuxièmement, nous avons présenté une mesure de similarité cross-linguistique des documents dans le but de récupérer et aligner automatiquement les documents comparables. Ensuite, nous avons proposé une méthode d’annotation cross-linguistique en termes de sentiments, afin d’étiqueter les documents source et cible avec des sentiments. Enfin, nous avons utilisé des mesures statistiques pour comparer l’accord des sentiments entre les documents comparables source et cible. Les méthodes présentées dans cette thèse ne dépendent pas d’une paire de langue bien déterminée, elles peuvent être appliquées sur toute autre couple de langue
The aim of this thesis is to study sentiments in comparable documents. First, we collect English, French and Arabic comparable corpora from Wikipedia and Euronews, and we align each corpus at the document level. We further gather English-Arabic news documents from local and foreign news agencies. The English documents are collected from BBC website and the Arabic documents are collected from Al-jazeera website. Second, we present a cross-lingual document similarity measure to automatically retrieve and align comparable documents. Then, we propose a cross-lingual sentiment annotation method to label source and target documents with sentiments. Finally, we use statistical measures to compare the agreement of sentiments in the source and the target pair of the comparable documents. The methods presented in this thesis are language independent and they can be applied on any language pair

Relations that hold between causes and their effects are fundamental for a wide range of different sectors. Automatically finding sentences that express such relations may for example be of great interest for the economy or political institutions. However, for many languages other than English, a lack of training resources for this task needs to be dealt with. In recent years, large, pretrained transformer-based model architectures have proven to be very effective for tasks involving cross-lingual transfer such as cross-lingual language inference, as well as multilingual named entity recognition, POS-tagging and dependency parsing, which may hint at similar potentials for causality detection. In this thesis, we define causality detection as a binary labelling problem and use cross-lingual transfer to alleviate data scarcity for German and Swedish by using three different classifiers that make either use of multilingual sentence embeddings obtained from a pretrained encoder or pretrained multilingual language models. The source languages in most of our experiments will be English, for Swedish we however also use a small German training set and a combination of English and German training data.  We try out zero-shot transfer as well as making use of limited amounts of target language data either as a development set or as additional training data in a few-shot setting. In the latter scenario, we explore the impact of varying sizes of training data. Moreover, the problem of data scarcity in our situation also makes it necessary to work with data from different annotation projects. We also explore how much this would impact our result. For German as a target language, our results in a zero-shot scenario expectedly fall short in comparison with monolingual experiments, but F1-macro scores between 60 and 65 in cases where annotation did not differ drastically still signal that it was possible to transfer at least some knowledge. When introducing only small amounts of target language data, already notable improvements were observed and with the full German training data of about 3,000 sentences combined with the most suitable English data set, the performance for German in some scenarios even almost matches the state of the art for monolingual experiments on English. The best zero-shot performance on the Swedish data was even outperforming the scores achieved for German. However, due to problems with the additional Swedish training data, we were not able to improve upon the zero-shot performance in a few-shot setting in a similar manner as it was the case for German.

Die Sprache der Judenfeindschaft in der frühen Neuzeit (1450–1700) steht im „Grenzbereich zwischen sprach- und kulturwissenschaftlicher Antisemitismusforschung“, die die „zeittypischen Argumentationsstrategien, Begründungsmuster und sprachlichen Darstellungsmittel in antijüdischen Schriften“ der Frühen Neuzeit in den Blick nimmt, um die „Zusammenhänge von vormodernem (religiösem) ,Antijudaismus‘ und modernem (rassistisch-säkularem) ,Antisemitismus‘“ zu erhellen (Klappentext). Methodisch ruht die Studie auf einer Verschränkung zwischen struktureller Semantik und einer die Argumentationsstrukturen hervorhebenden Textlinguistik auf. Sie knüpft damit unmittelbar an die Vorarbeiten von Nicoline Hortzitz an, die in der Sprache der Judenfeindschaft die Ergebnisse ihrer Dissertation zum Früh-Antisemitismus (1988), der Monographie Judenarzt (1994) und diverser Aufsätze in einer „Gesamtbewertung der Befunde“ (S. V) vorlegt. In seiner Besprechung der Dissertation zum Früh-Antisemitismus merkte Matthias Richter kritisch an, daß „infolge des strukturalistischen Ansatzes die individuellen Differenzen zwischen den untersuchten Autoren zu sehr vernachlässigt werden.“¹ Die „minuziöse[] und aufwendige[] Systematisierung des sprachlichen Instrumentariums der Früh-Antisemiten“² setzt die vorliegende Monographie in der Ausweitung des Untersuchungszeitraums fort, ohne dabei jedoch eine „strukturelle Untersuchung“ liefern zu wollen, wie die Akzentverschiebung im Titel nahe zu legen scheint.

Pesquisas recentes em medicina genômica sugerem que fatores de risco que incidem desde a concepção de uma criança até o final de sua adolescência podem influenciar no desenvolvimento de doenças crônicas da idade adulta. Artigos científicos com descobertas e estudos inovadores sobre o tema indicam que a epigenética deve ser explorada para prevenir doenças de alta prevalência como doenças cardiovasculares, diabetes e obesidade. A grande quantidade de artigos disponibilizados diariamente dificulta a atualização de profissionais, uma vez que buscas por informação exata se tornam complexas e dispendiosas em relação ao tempo gasto na procura e análise dos resultados. Algumas tecnologias e técnicas computacionais podem apoiar a manipulação dos grandes repositórios de informações biomédicas, assim como a geração de conhecimento. O presente trabalho pesquisa a descoberta automática de artigos científicos que relacionem doenças crônicas e fatores de risco para as mesmas em registros clínicos de pacientes. Este trabalho também apresenta o desenvolvimento de um arcabouço de software para sistemas de vigilância que alertem profissionais de saúde sobre problemas no desenvolvimento humano. A efetiva transformação dos resultados de pesquisas biomédicas em conhecimento possível de ser utilizado para beneficiar a saúde pública tem sido considerada um domínio importante da informática. Este domínio é denominado Bioinformática Translacional (BUTTE,2008). Considerando-se que doenças crônicas são, mundialmente, um problema sério de saúde e lideram as causas de mortalidade com 60% de todas as mortes, o presente trabalho poderá possibilitar o uso direto dos resultados dessas pesquisas na saúde pública e pode ser considerado um trabalho de Bioinformática Translacional.
Genomic medicine has suggested that the exposure to risk factors since conception may influence gene expression and consequently induce the development of chronic diseases in adulthood. Scientific papers bringing up these discoveries indicate that epigenetics must be exploited to prevent diseases of high prevalence, such as cardiovascular diseases, diabetes and obesity. A large amount of scientific information burdens health care professionals interested in being updated, once searches for accurate information become complex and expensive. Some computational techniques might support management of large biomedical information repositories and discovery of knowledge. This study presents a framework to support surveillance systems to alert health professionals about human development problems, retrieving scientific papers that relate chronic diseases to risk factors detected on a patient\'s clinical record. As a contribution, healthcare professionals will be able to create a routine with the family, setting up the best growing conditions. According to Butte, the effective transformation of results from biomedical research into knowledge that actually improves public health has been considered an important domain of informatics and has been called Translational Bioinformatics. Since chronic diseases are a serious health problem worldwide and leads the causes of mortality with 60% of all deaths, this scientific investigation will probably enable results from bioinformatics researches to directly benefit public health.

This thesis explores to what extent Multinomial Naive Bayes (MNB) and Support Vector Machines (SVM) classifiers can be used to determine the polarity of news, specifically the news coverage of Sweden by the Russian state-funded news outlets RT and Sputnik. Three experiments are conducted.  In the first experiment, an MNB and an SVM classifier are trained with the Large Movie Review Dataset (Maas et al., 2011) with a varying number of samples to determine how training data size affects classifier performance.  In the second experiment, the classifiers are trained with 300 positive, negative, and neutral news articles (Agarwal et al., 2019) and tested on 95 RT and Sputnik news articles about Sweden (Bengtsson, 2019) to determine if the domain specificity of the training data outweighs its limited size.  In the third experiment, the movie-trained classifiers are put up against the domain-specific classifiers to determine if well-trained classifiers from another domain perform better than relatively untrained, domain-specific classifiers.  Four different types of feature sets (unigrams, unigrams without stop words removal, bigrams, trigrams) were used in the experiments. Some of the model parameters (TF-IDF vs. feature count and SVM’s C parameter) were optimized with 10-fold cross-validation.  Other than the superior performance of SVM, the results highlight the need for comprehensive and domain-specific training data when conducting machine learning tasks, as well as the benefits of feature engineering, and to a limited extent, the removal of stop words. Interestingly, the classifiers performed the best on the negative news articles, which made up most of the test set (and possibly of Russian news coverage of Sweden in general).

Prvotním cílem tohoto projektu bylo prostudovat problematiku jazykového modelování pro rozpoznávání řeči a techniky pro získávání textových dat z Webu. Text představuje základní techniky rozpoznávání řeči a detailněji popisuje jazykové modely založené na statistických metodách. Zvláště se práce zabývá kriterii pro vyhodnocení kvality jazykových modelů a systémů pro rozpoznávání řeči. Text dále popisuje modely a techniky dolování dat, zvláště vyhledávání informací. Dále jsou představeny problémy spojené se získávání dat z webu, a v kontrastu s tím je představen vyhledávač Google. Součástí projektu byl návrh a implementace systému pro získávání textu z webu, jehož detailnímu popisu je věnována náležitá pozornost. Nicméně, hlavním cílem práce bylo ověřit, zda data získaná z Webu mohou mít nějaký přínos pro rozpoznávání řeči. Popsané techniky se tak snaží najít optimální způsob, jak data získaná z Webu použít pro zlepšení ukázkových jazykových modelů, ale i modelů nasazených v reálných rozpoznávacích systémech.

碩士
國立清華大學
資訊系統與應用研究所
106
Recently, the data-driven machine learning approaches have shown their successes on many text classification tasks for a resource-abundant language. However, there are still many languages that lack of sufficient enough labeled data for carrying out the same specific tasks. They may be costly to obtain high-quality parallel corpus or cannot rely on automated machine translation due to unreliable or unavailable machine translation tools in those low-resource languages. In this work, we propose an effective transfer learning method in the scenarios where the large-scale cross-lingual data is not available. It combines transfer learning schemes of parameter sharing (parameter based) and domain adaptation (feature based) that are joint trained with high-resource and low-resource languages together. We conducted the cross-lingual transfer learning experiments on text classification on sentiment, subjectivity and question types from English to Chinese and from English to Vietnamese respectively. The experiments show that the proposed approach significantly outperformed the state-of-the-art models that are trained merely with monolingual data on the corresponding benchmarks.

碩士
國立中山大學
資訊管理學系研究所
92
With the emergence and proliferation of Internet services and e-commerce applications, a tremendous amount of information is accessible online, typically as textual documents. To facilitate subsequent access to and leverage from this information, the efficient and effective management—specifically, text categorization—of the ever-increasing volume of textual documents is essential to organizations and person. Existing text categorization techniques focus mainly on categorizing monolingual documents. However, with the globalization of business environments and advances in Internet technology, an organization or person often retrieves and archives documents in different languages, thus creating the need for cross-lingual text categorization. Motivated by the significance of and need for such a cross-lingual text categorization technique, this thesis designs a technique with two different category assignment methods, namely, individual- and cluster-based. The empirical evaluation results show that the cross-lingual text categorization technique performs well and the cluster-based method outperforms the individual-based method.

Law Yin Yee.
Thesis (M.Phil.)--Chinese University of Hong Kong, 2003.
Includes bibliographical references (leaves 130-134).
Abstracts in English and Chinese.
Chapter 1 --- Introduction --- p.1
Chapter 1.1 --- Cross-Lingual Information Retrieval (CLIR) --- p.2
Chapter 1.2 --- Bilingual Term Association Mining --- p.5
Chapter 1.3 --- Our Contributions --- p.6
Chapter 1.3.1 --- CLIR --- p.6
Chapter 1.3.2 --- Bilingual Term Association Mining --- p.7
Chapter 1.4 --- Thesis Organization --- p.8
Chapter 2 --- Related Work --- p.9
Chapter 2.1 --- CLIR Techniques --- p.9
Chapter 2.1.1 --- Existing Approaches --- p.9
Chapter 2.1.2 --- Difference Between Our Model and Existing Approaches --- p.13
Chapter 2.2 --- Bilingual Term Association Mining Techniques --- p.13
Chapter 2.2.1 --- Existing Approaches --- p.13
Chapter 2.2.2 --- Difference Between Our Model and Existing Approaches --- p.17
Chapter 3 --- Cross-Lingual Information Retrieval (CLIR) --- p.18
Chapter 3.1 --- Cross-Lingual Query Processing and Translation --- p.18
Chapter 3.1.1 --- Query Context and Document Context Generation --- p.20
Chapter 3.1.2 --- Context-Based Query Translation --- p.23
Chapter 3.1.3 --- Query Term Weighting --- p.28
Chapter 3.1.4 --- Final Weight Calculation --- p.30
Chapter 3.2 --- Retrieval on Documents and Automated Summaries --- p.32
Chapter 4 --- Experiments on Cross-Lingual Information Retrieval --- p.38
Chapter 4.1 --- Experimental Setup --- p.38
Chapter 4.2 --- Results of English-to-Chinese Retrieval --- p.45
Chapter 4.2.1 --- Using Mono-Lingual Retrieval as the Gold Standard --- p.45
Chapter 4.2.2 --- Using Human Relevance Judgments as the Gold Stan- dard --- p.49
Chapter 4.3 --- Results of Chinese-to-English Retrieval --- p.53
Chapter 4.3.1 --- Using Mono-lingual Retrieval as the Gold Standard --- p.53
Chapter 4.3.2 --- Using Human Relevance Judgments as the Gold Stan- dard --- p.57
Chapter 5 --- Discovering Comparable Multi-lingual Online News for Text Mining --- p.61
Chapter 5.1 --- Story Representation --- p.62
Chapter 5.2 --- Gloss Translation --- p.64
Chapter 5.3 --- Comparable News Discovery --- p.67
Chapter 6 --- Mining Bilingual Term Association Based on Co-occurrence --- p.75
Chapter 6.1 --- Bilingual Term Cognate Generation --- p.75
Chapter 6.2 --- Term Mining Algorithm --- p.77
Chapter 7 --- Phonetic Matching --- p.87
Chapter 7.1 --- Algorithm Design --- p.87
Chapter 7.2 --- Discovering Associations of English Terms and Chinese Terms --- p.93
Chapter 7.2.1 --- Converting English Terms into Phonetic Representation --- p.93
Chapter 7.2.2 --- Discovering Associations of English Terms and Man- darin Chinese Terms --- p.100
Chapter 7.2.3 --- Discovering Associations of English Terms and Can- tonese Chinese Terms --- p.104
Chapter 8 --- Experiments on Bilingual Term Association Mining --- p.111
Chapter 8.1 --- Experimental Setup --- p.111
Chapter 8.2 --- Result and Discussion of Bilingual Term Association Mining Based on Co-occurrence --- p.114
Chapter 8.3 --- Result and Discussion of Phonetic Matching --- p.121
Chapter 9 --- Conclusions and Future Work --- p.126
Chapter 9.1 --- Conclusions --- p.126
Chapter 9.1.1 --- CLIR --- p.126
Chapter 9.1.2 --- Bilingual Term Association Mining --- p.127
Chapter 9.2 --- Future Work --- p.128
Bibliography --- p.134
Chapter A --- Original English Queries --- p.135
Chapter B --- Manual translated Chinese Queries --- p.137
Chapter C --- Pronunciation symbols used by the PRONLEX Lexicon --- p.139
Chapter D --- Initial Letter-to-Phoneme Tags --- p.141
Chapter E --- English Sounds with their Chinese Equivalents --- p.143

碩士
國立中山大學
資訊管理學系研究所
93
Text categorization deals with the automatic learning of a text categorization model from a training set of preclassified documents on the basis of their contents and the assignment of unclassified documents to appropriate categories. Most of existing text categorization techniques deal with monolingual documents (i.e., all documents are written in one language) during the text categorization model learning and category assignment (or prediction). However, with the globalization of business environments and advances in Internet technology, an organization or individual often generates/acquires and subsequently archives documents in different languages, thus creating the need for cross-lingual text categorization (CLTC). Existing studies on CLTC focus on the prediction-corpus translation-based approach that lacks of a systematic mechanism for reducing translation noises; thus, limiting their cross-lingual categorization effectiveness. Motivated by the needs of providing more effective CLTC support, we design a training-corpus translation-based CLTC approach. Using the prediction-corpus translation-based approach as the performance benchmark, our empirical evaluation results show that our proposed CLTC approach achieves significantly better classification effectiveness than the benchmark approach does in both Chinese

Identifying sentiment in a low-resource language is essential for understanding opinions internationally and for responding to the urgent needs of locals affected by disaster incidents in different world regions. While tools and resources for recognizing sentiment in high-resource languages are plentiful, determining the most effective methods for achieving this task in a low-resource language which lacks annotated data is still an open research question. Most existing approaches for cross-lingual sentiment analysis to date have relied on high-resource machine translation systems, large amounts of parallel data, or resources only available for Indo-European languages. This work presents methods, resources, and strategies for identifying sentiment cross-lingually in a low-resource language. We introduce a cross-lingual sentiment model which can be trained on a high-resource language and applied directly to a low-resource language. The model offers the feature of lexicalizing the training data using a bilingual dictionary, but can perform well without any translation into the target language. Through an extensive experimental analysis, evaluated on 17 target languages, we show that the model performs well with bilingual word vectors pre-trained on an appropriate translation corpus. We compare in-genre and in-domain parallel corpora, out-of-domain parallel corpora, in-domain comparable corpora, and monolingual corpora, and show that a relatively small, in-domain parallel corpus works best as a transfer medium if it is available. We describe the conditions under which other resources and embedding generation methods are successful, and these include our strategies for leveraging in-domain comparable corpora for cross-lingual sentiment analysis. To enhance the ability of the cross-lingual model to identify sentiment in the target language, we present new feature representations for sentiment analysis that are incorporated in the cross-lingual model: bilingual sentiment embeddings that are used to create bilingual sentiment scores, and a method for updating the sentiment embeddings during training by lexicalization of the target language. This feature configuration works best for the largest number of target languages in both untargeted and targeted cross-lingual sentiment experiments. The cross-lingual model is studied further by evaluating the role of the source language, which has traditionally been assumed to be English. We build cross-lingual models using 15 source languages, including two non-European and non-Indo-European source languages: Arabic and Chinese. We show that language families play an important role in the performance of the model, as does the morphological complexity of the source language. In the last part of the work, we focus on sentiment analysis towards targets. We study Arabic as a representative morphologically complex language and develop models and morphological representation features for identifying entity targets and sentiment expressed towards them in Arabic open-domain text. Finally, we adapt our cross-lingual sentiment models for the detection of sentiment towards targets. Through cross-lingual experiments on Arabic and English, we demonstrate that our findings regarding resources, features, and language also hold true for the transfer of targeted sentiment.

碩士
慈濟大學
醫學資訊學系碩士班
107
People often use the Internet to find answers of healthcare questions. Many healthcare information websites thus collect and maintain a database of frequently asked questions (FAQs) answered by healthcare professionals. However, with the increasing amount of FAQs, users are difficult to identify specific FAQs to satisfy their information needs, and moreover many reliable healthcare FAQs are written in English. Therefore, we propose a technique to rank English healthcare FAQs with respect to Chinese healthcare questions. The technique considers information need aspects, which indicate basic types of healthcare information required by people. By recognizing the aspects, our technique can improve the performance of various kinds of FAQ retrievers. Empirical evaluation on thousands of English and Chinese healthcare FAQs show that our technique can significantly enhance several kinds of FAQ retrievers. Our technique can thus help the users to find reliable answers of healthcare questions.