Many projects exist whose purpose is to augment raw data with annotations that increase the usefulness of the data. The number of these projects is rapidly growing and in the age of “big data” the amount of data to be annotated is likewise growing within each project. One common use of such data is in supervised machine learning, which requires labeled data to train a predictive model. Annotation is often a very expensive proposition, particularly for structured data. The purpose of this dissertation is to explore methods of reducing the cost of creating such data sets, including annotated text corpora.We focus on active learning to address the annotation problem. Active learning employs models trained using machine learning to identify instances in the data that are most informative and least costly. We introduce novel techniques for adapting vanilla active learning to situations wherein data instances are of varying benefit and cost, annotators request work “on-demand,” and there are multiple, fallible annotators of differing levels of accuracy and cost. In order to account for data instances of varying cost, we build a model of cost from real annotation data based on a user study. We also introduce a novel cost-conscious active learning algorithm which we call return-on-investment, that selects instances for annotation that contain the most benefit per unit cost. To address the issue of annotators that request instances “on-demand,” we develop a parallel, “no-wait” framework that performs computation while the annotator is annotating. As a result, annotators need not wait for the computer to determine the best instance for them to annotate—a common problem with existing approaches. Finally, we introduce a Bayesian model designed to simultaneously infer ground truth annotations from noisy annotations, infer each individual annotators accuracy, and predict its own accuracy on unseen data, without the use of a held-out set. We extend ROI-based active learning and our annotation framework to handle multiple annotators using this model. As a whole, our work shows that the techniques introduced in this dissertation reduce the cost of annotation in scenarios that are more true-to-life than previous research.

[pt] Rotular as classes gramaticais ao longo de uma sentença - part-ofspeech tagging - é uma das primeiras tarefas de processamento de linguagem natural, fornecendo atributos importantes para realizar tarefas de alta complexidade. A representação de texto a nível de palavra tem sido amplamente adotada, tanto através de uma codificação esparsa convencional, e.g. bagofwords; quanto por uma representação distribuída, como os sofisticados modelos de word-embedding usados para descrever informações sintáticas e semânticas. Um problema importante desse tipo de codificação é a carência de aspectos morfológicos. Além disso, os sistemas atuais apresentam uma precisão por token em torno de 97 por cento. Contudo, quando avaliados por sentença, apresentam um resultado mais modesto com uma taxa de acerto em torno de 55−57 por cento. Neste trabalho, nós demonstramos como utilizar n-grams para derivar automaticamente atributos esparsos e morfológicos para processamento de texto. Essa representação permite que redes neurais realizem a tarefa de POS-Tagging a partir de uma representação a nível de caractere. Além disso, introduzimos uma estratégia de regularização capaz de selecionar atributos específicos para cada neurônio. A utilização de regularização embutida em nossos modelos produz duas variantes. A primeira compartilha os n-grams selecionados globalmente entre todos os neurônios de uma camada; enquanto que a segunda opera uma seleção individual para cada neurônio, de forma que cada neurônio é sensível apenas aos n-grams que mais o estimulam. Utilizando a abordagem apresentada, nós geramos uma alta quantidade de características que representam afeições morfossintáticas relevantes baseadas a nível de caractere. Nosso POS tagger atinge a acurácia de 96, 67 por cento no corpus Mac-Morpho para o Português. / [en] Part-of-speech tagging is one of the primary stages in natural language processing, providing useful features for performing higher complexity tasks. Word level representations have been largely adopted, either through a conventional sparse codification, such as bag-of-words, or through a distributed representation, like the sophisticated word embedded models used to describe syntactic and semantic information. A central issue on these codifications is the lack of morphological aspects. In addition, recent taggers present per-token accuracies around 97 percent. However, when using a persentence metric, the good taggers show modest accuracies, scoring around 55-57 percent. In this work, we demonstrate how to use n-grams to automatically derive morphological sparse features for text processing. This representation allows neural networks to perform POS tagging from a character-level input. Additionally, we introduce a regularization strategy capable of selecting specific features for each layer unit. As a result, regarding n-grams selection, using the embedded regularization in our models produces two variants. The first one shares globally selected features among all layer units, whereas the second operates individual selections for each layer unit, so that each unit is sensible only to the n-grams that better stimulate it. Using the proposed approach, we generate a high number of features which represent relevant morphosyntactic affection based on a character-level input. Our POS tagger achieves the accuracy of 96.67 percent in the Mac-Morpho corpus for Portuguese.

No description available.

Cross-lingual transfer learning has been the driving force of low-resource natural language processing in recent years, relying on massively multilingual language models with hopes of solving the data scarcity issue for languages with a limited digital presence. However, this "one-size-fits-all" approach is not equally applicable to all low-resource languages, suggesting limitations of such models in cross-lingual transfer. Besides, known similarities and phylogenetic relationships between source and target languages are often overlooked. In this work, the emphasis is placed on Faroese, a low-resource North Germanic language with several closely related resource-rich sibling languages. The cross-lingual transfer potential from these strong Scandinavian source candidates, as well as from additional genetically related, geographically proximate, and syntactically similar source languages is studied in single-source and multi-source experiments, in terms of Faroese syntactic parsing and part-of-speech tagging. In addition, the effect of task-specific fine-tuning on monolingual, linguistically informed smaller multilingual, and massively multilingual pre-trained language models is explored. The results suggest Icelandic as a strong source candidate, however, only when fine-tuning a monolingual model. With multilingual models, task-specific fine-tuning in Norwegian and Swedish seems even more beneficial. Although they do not surpass fully Scandinavian fine-tuning, models trained on genetically related and syntactically similar languages produce good results. Additionally, the findings indicate that multilingual models outperform models pre-trained on a single language, and that even better results can be achieved using a smaller, linguistically informed model, compared to a massively multilingual one.

[pt] Part-of-speech (POS) tagging é o processo de categorizar cada palavra de uma sentença com sua devida classe morfossintática (verbo, substantivo, adjetivo e etc). POS tagging é considerada uma atividade fundamental no processo de construção de aplicações de processamento de linguagem natural (PLN), muitas dessas aplicações, em algum ponto, demandam esse tipo de informação. Nesse trabalho, construímos um POS tagger para o Português Contemporâneo e o Português Histórico, baseado em uma arquitetura de rede neural recorrente. Tradicionalmente a construção dessas ferramentas requer muitas features específicas do domínio da linguagem e dados externos ao conjunto de treino, mas nosso POS tagger não usa esses requisitos. Treinamos uma rede Bidirectional Long short-term memory (BLSTM), que se beneficia das representações de word embeddings e character embeddings das palavras, para atividade de classificação morfossintática. Testamos nosso POS tagger em três corpora diferentes: a versão original do corpus MacMorpho, a versão revisada do corpus Mac-Morpho e no corpus Tycho Brahe. Nós obtemos um desempenho ligeiramente melhor que os sistemas estado da arte nos três corpora: 97.83 por cento de acurácia para o Mac-Morpho original, 97.65 por cento de acurácia para o Mac-Morpho revisado e 97.35 por cento de acurácia para Tycho Brahe. Conseguimos, também, uma melhora nos três corpora para a medida de acurácia fora do vocabulário, uma acurácia especial calculada somente sobre as palavras desconhecidas do conjunto de treino. Realizamos ainda um estudo comparativo para verificar qual dentre os mais populares algoritmos de criação de word embedding (Word2Vec, FastText, Wang2Vec e Glove), é mais adequado para a atividade POS tagging em Português. O modelo de Wang2Vec mostrou um desempenho superior. / [en] Part-of-speech (POS) tagging is a process of labeling each word in a sentence with a morphosyntactic class (verb, noun, adjective and etc). POS tagging is a fundamental part of the linguistic pipeline, most natural language processing (NLP) applications demand, at some step, part-of-speech information. In this work, we constructed a POS tagger for Contemporary Portuguese and Historical Portuguese, using a recurrent neural network architecture. Traditionally the development of these tools requires many handcraft features and external data, our POS tagger does not use these elements. We trained a Bidirectional Long short-term memory (BLSTM) network that benefits from the word embeddings and character embeddings representations of the words, for morphosyntactic classification. We tested our POS tagger on three different corpora: the original version of the Mac-Morpho corpus, the revised version of the Mac-Morpho corpus, and the Tycho Brahe corpus. We produce state-of-the-art POS taggers for the three corpora: 97.83 percent accuracy on the original Mac-Morpho corpus, 97.65 percent accuracy on the revised Mac-Morpho and 97.35 percent accuracy on the Tycho Brahe corpus. We also achieved an improvement in the three corpora in out-of-vocabulary accuracy, that is the accuracy on words not seen in training sentences. We also performed a comparative study to test which different types of word embeddings (Word2Vec, FastText, Wang2Vec, and Glove) is more suitable for Portuguese POS tagging. The Wang2Vec model showed higher performance.

Contemporary approaches to natural language processing are predominantly based on statistical machine learning from large amounts of text, which has been manually annotated with the linguistic structure of interest. However, such complete supervision is currently only available for the world's major languages, in a limited number of domains and for a limited range of tasks. As an alternative, this dissertation considers methods for linguistic structure prediction that can make use of incomplete and cross-lingual supervision, with the prospect of making linguistic processing tools more widely available at a lower cost. An overarching theme of this work is the use of structured discriminative latent variable models for learning with indirect and ambiguous supervision; as instantiated, these models admit rich model features while retaining efficient learning and inference properties. The first contribution to this end is a latent-variable model for fine-grained sentiment analysis with coarse-grained indirect supervision. The second is a model for cross-lingual word-cluster induction and the application thereof to cross-lingual model transfer. The third is a method for adapting multi-source discriminative cross-lingual transfer models to target languages, by means of typologically informed selective parameter sharing. The fourth is an ambiguity-aware self- and ensemble-training algorithm, which is applied to target language adaptation and relexicalization of delexicalized cross-lingual transfer parsers. The fifth is a set of sequence-labeling models that combine constraints at the level of tokens and types, and an instantiation of these models for part-of-speech tagging with incomplete cross-lingual and crowdsourced supervision. In addition to these contributions, comprehensive overviews are provided of structured prediction with no or incomplete supervision, as well as of learning in the multilingual and cross-lingual settings. Through careful empirical evaluation, it is established that the proposed methods can be used to create substantially more accurate tools for linguistic processing, compared to both unsupervised methods and to recently proposed cross-lingual methods. The empirical support for this claim is particularly strong in the latter case; our models for syntactic dependency parsing and part-of-speech tagging achieve the hitherto best published results for a wide number of target languages, in the setting where no annotated training data is available in the target language.

A family of Natural Language Processing (NLP) tasks such as part-of- speech (PoS) tagging, Named Entity Recognition (NER), and Multiword Expression (MWE) identification all involve assigning labels to sequences of words in text (sequence labeling). Most modern machine learning approaches to sequence labeling utilize word embeddings, learned representations of text, in which words with similar meanings have similar representations. Quite recently, contextualized word embeddings have garnered much attention because, unlike pretrained context- insensitive embeddings such as word2vec, they are able to capture word meaning in context. In this thesis, I evaluate the performance of different embedding setups (context-sensitive, context-insensitive word, as well as task-specific word, character, lemma, and PoS) on the three abovementioned sequence labeling tasks using a deep learning model (BiLSTM) and Portuguese datasets. v

No description available.

Contemporary approaches to natural language processing are predominantly based on statistical machine learning from large amounts of text, which has been manually annotated with the linguistic structure of interest. However, such complete supervision is currently only available for the world's major languages, in a limited number of domains and for a limited range of tasks. As an alternative, this dissertation considers methods for linguistic structure prediction that can make use of incomplete and cross-lingual supervision, with the prospect of making linguistic processing tools more widely available at a lower cost. An overarching theme of this work is the use of structured discriminative latent variable models for learning with indirect and ambiguous supervision; as instantiated, these models admit rich model features while retaining efficient learning and inference properties. The first contribution to this end is a latent-variable model for fine-grained sentiment analysis with coarse-grained indirect supervision. The second is a model for cross-lingual word-cluster induction and the application thereof to cross-lingual model transfer. The third is a method for adapting multi-source discriminative cross-lingual transfer models to target languages, by means of typologically informed selective parameter sharing. The fourth is an ambiguity-aware self- and ensemble-training algorithm, which is applied to target language adaptation and relexicalization of delexicalized cross-lingual transfer parsers. The fifth is a set of sequence-labeling models that combine constraints at the level of tokens and types, and an instantiation of these models for part-of-speech tagging with incomplete cross-lingual and crowdsourced supervision. In addition to these contributions, comprehensive overviews are provided of structured prediction with no or incomplete supervision, as well as of learning in the multilingual and cross-lingual settings. Through careful empirical evaluation, it is established that the proposed methods can be used to create substantially more accurate tools for linguistic processing, compared to both unsupervised methods and to recently proposed cross-lingual methods. The empirical support for this claim is particularly strong in the latter case; our models for syntactic dependency parsing and part-of-speech tagging achieve the hitherto best published results for a wide number of target languages, in the setting where no annotated training data is available in the target language.

In the Commedia of Dante, a poem 14,233 lines in length, some 7,500 words occur only once. These are the hapax. Fewer than 2% of these constitute a minute but distinct subset—the hapax for which there are one or more words in the poem whose spelling is identical but whose meaning is different. These are what I call same-spelling hapax. I identify four categories: part-of-speech, homograph, locus, and name. Analysis of the same-spelling hapax illuminates a poetic strategy continuously in use throughout the poem. This is to use the one-word overlap of Rhyme and line number. Not only is it highly probable that a same-spelling hapax will be a rhyme-word, but it is also probable that it will occupy a rhyme-word’s most significant position—the one place—the single word—where the two intertwined formal entities that shape each canto coincide. Every three lines, their tension-resolving this-word-only union intensifies the reader’s attention and understanding alike.