Narrativas clínicas são normalmente escritas em linguagem natural devido a seu poder descritivo e facilidade de comunicação entre os especialistas. Processar esses dados para fins de descoberta de conhecimento e coleta de estatísticas exige técnicas de extração de informações, com alguns resultados já apresentados na literatura para o domínio jornalístico, mas ainda raras no domínio médico. O presente trabalho visa desenvolver um classificador de laudos de anatomia patológica que seja capaz de inferir a topografia e a morfologia de um câncer na Classificação Internacional de Doenças para Oncologia (CID-O). Dados fornecidos pelo A.C. Camargo Cancer Center em São Paulo foram utilizados para treinamento e validação. Técnicas de processamento de linguagem natural (PLN) aliadas a classificadores bayesianos foram exploradas na busca de qualidade da recuperação da informação, avaliada por meio da medida-F2. Valores acima de 74% para o grupo topográfico e de 61% para o grupo morfológico são relatados, com pequena contribuição das técnicas de PLN e suavização. Os resultados corroboram trabalhos similares e demonstram a necessidade de retreinamento das ferramentas de PLN no domínio médico. / Clinical reports are usually written in natural language due to its descriptive power and ease of communication among specialists. Processing data for knowledge discovery and statistical analysis requires information retrieval techniques, already established for newswire texts, but still rare in the medical subdomain. The present work aims at developing an automated classifier of pathology reports, which should be able to infer the topography and the morphology classes of a cancer using codes of the International Classification of Diseases for Oncology (ICD-O). Data provided by the A.C. Camargo Cancer Center located in Sao Paulo was used for training and validation. Techniques of natural language processing (NLP) and Bayes classifiers were used in search for information retrieval quality, evaluated by F2-score. Measures upper than 74% in the topographic group and 61% in the morphologic group are reported, with small contribution from NLP or smoothing techniques. The results agree with similar studies and show that a retraining of NLP tools in the medical domain is necessary.

El lenguaje figurado representa una de las tareas más difíciles del procesamiento del lenguaje natural. A diferencia del lenguaje literal, el lenguaje figurado hace uso de recursos lingüísticos tales como la ironía, el humor, el sarcasmo, la metáfora, la analogía, entre otros, para comunicar significados indirectos que la mayoría de las veces no son interpretables sólo en términos de información sintáctica o semántica. Por el contrario, el lenguaje figurado refleja patrones del pensamiento que adquieren significado pleno en contextos comunicativos y sociales, lo cual hace que tanto su representación lingüística, así como su procesamiento computacional, se vuelvan tareas por demás complejas. En este contexto, en esta tesis de doctorado se aborda una problemática relacionada con el procesamiento del lenguaje figurado a partir de patrones lingüísticos. En particular, nuestros esfuerzos se centran en la creación de un sistema capaz de detectar automáticamente instancias de humor e ironía en textos extraídos de medios sociales. Nuestra hipótesis principal se basa en la premisa de que el lenguaje refleja patrones de conceptualización; es decir, al estudiar el lenguaje, estudiamos tales patrones. Por tanto, al analizar estos dos dominios del lenguaje figurado, pretendemos dar argumentos respecto a cómo la gente los concibe, y sobre todo, a cómo esa concepción hace que tanto humor como ironía sean verbalizados de una forma particular en diversos medios sociales. En este contexto, uno de nuestros mayores intereses es demostrar cómo el conocimiento que proviene del análisis de diferentes niveles de estudio lingüístico puede representar un conjunto de patrones relevantes para identificar automáticamente usos figurados del lenguaje. Cabe destacar que contrario a la mayoría de aproximaciones que se han enfocado en el estudio del lenguaje figurado, en nuestra investigación no buscamos dar argumentos basados únicamente en ejemplos prototípicos, sino en textos cuyas características / Reyes Pérez, A. (2012). Linguistic-based Patterns for Figurative Language Processing: The Case of Humor Recognition and Irony Detection [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/16692 / Palancia

[pt] Aprendizado de estruturas consiste em aprender um mapeamento de variáveis de entrada para saídas estruturadas a partir de exemplos de pares entrada-saída. Vários problemas importantes podem ser modelados desta maneira. O processamento de linguagem natural provê diversas tarefas que podem ser formuladas e solucionadas através do aprendizado de estruturas. Por exemplo, parsing de dependência envolve o reconhecimento de uma árvore implícita em uma frase. Geração de atributos é uma sub-tarefa importante do aprendizado de estruturas. Geralmente, esta sub-tarefa é realizada por um especialista que constrói gabaritos de atributos complexos e discriminativos através da combinação dos atributos básicos disponíveis na entrada. Esta é uma forma limitada e cara para geração de atributos e é reconhecida como um gargalo de modelagem. Neste trabalho, propomos um método automático para geração de atributos para problemas de aprendizado de estruturas. Este método é guiado por entropia já que é baseado na entropia condicional de variáveis locais de saída dados os atributos básicos. Comparamos experimentalmente o método proposto com dois métodos alternativos para geração de atributos: geração manual e métodos de kernel polinomial. Nossos resultados mostram que o método de geração de atributos guiado por entropia é superior aos dois métodos alternativos em diferentes aspectos. Nosso método é muito mais barato do que o método manual e computacionalmente mais rápido que o método baseado em kernel. Adicionalmente, ele permite o controle do seu poder de generalização mais facilmente do que métodos de kernel. Nós avaliamos nosso método em nove datasets envolvendo cinco tarefas de linguística computacional e quatro idiomas. Os sistemas desenvolvidos apresentam resultados comparáveis aos melhores sistemas atualmente e, particularmente para etiquetagem morfossintática, identificação de sintagmas, extração de citações e resolução de coreferência, obtêm os melhores resultados conhecidos para diferentes idiomas como Árabe, Chinês, Inglês e Português. Adicionalmente, nosso sistema de resolução de coreferência obteve o primeiro lugar na competição Conference on Computational Natural Language Learning 2012 Shared Task. O sistema vencedor foi determinado pela média de desempenho em três idiomas: Árabe, Chinês e Inglês. Nosso sistema obteve o melhor desempenho nos três idiomas avaliados. Nosso método de geração de atributos estende naturalmente o framework de aprendizado de estruturas e não está restrito a tarefas de processamento de linguagem natural. / [en] Structure learning consists in learning a mapping from inputs to structured outputs by means of a sample of correct input-output pairs. Many important problems fit into this setting. Natural language processing provides several tasks that can be formulated and solved as structure learning problems. Dependency parsing, for instance, involves the prediction of a tree underlying a sentence. Feature generation is an important subtask of structure learning which, usually, is partially solved by a domain expert that builds complex discriminative feature templates by conjoining the available basic features. This is a limited and expensive way to generate features and is recognized as a modeling bottleneck. In this work, we propose an automatic feature generation method for structure learning problems. This method is entropy guided since it generates complex features based on the conditional entropy of local output variables given the available input features. We experimentally compare the proposed method with two important alternative feature generation methods, namely manual template generation and polynomial kernel methods. Our experimental findings indicate that the proposed method is more attractive than both alternatives. It is much cheaper than manual templates and computationally faster than kernel methods. Additionally, it is simpler to control its generalization performance than with kernel methods. We evaluate our method on nine datasets involving five natural language processing tasks and four languages. The resulting systems present state-of-the-art comparable performances and, particularly on part-of-speech tagging, text chunking, quotation extraction and coreference resolution, remarkably achieve the best known performances on different languages like Arabic, Chinese, English, and Portuguese. Furthermore, our coreference resolution systems achieve the very first place on the Conference on Computational Natural Language Learning 2012 Shared Task. The competing systems were ranked by the mean score over three languages: Arabic, Chinese and English. Our approach obtained the best performances among all competitors for all the three languages. Our feature generation method naturally extends the general structure learning framework and is not restricted to natural language processing tasks.

abstract: Smart home assistants are becoming a norm due to their ease-of-use. They employ spoken language as an interface, facilitating easy interaction with their users. Even with their obvious advantages, natural-language based interfaces are not prevalent outside the domain of home assistants. It is hard to adopt them for computer-controlled systems due to the numerous complexities involved with their implementation in varying fields. The main challenge is the grounding of natural language base terms into the underlying system's primitives. The existing systems that do use natural language interfaces are specific to one problem domain only. In this thesis, a domain-agnostic framework that creates natural language interfaces for computer-controlled systems has been developed by making the mapping between the language constructs and the system primitives customizable. The framework employs ontologies built using OWL (Web Ontology Language) for knowledge representation purposes and machine learning models for language processing tasks. It has been evaluated within a simulation environment consisting of objects and a robot. This environment has been deployed as a web application, providing anonymous user testing for evaluation, and generating training data for machine learning components. Performance evaluation has been done on metrics such as time taken for a task or the number of instructions given by the user to the robot to accomplish a task. Additionally, the framework has been used to create a natural language interface for a database system to demonstrate its domain independence. / Dissertation/Thesis / Masters Thesis Software Engineering 2020

La thèse vise à explorer des modèles et algorithmes d'extraction de connaissance et d'interconnexion de bases de données hétérogènes, appliquée à la gestion de contenus tels que rencontrés fréquemment dans le quotidien des journalistes. Le travail se déroulera dans le cadre du projet ANR ContentCheck (2016-2019) qui fournit le financement et dans le cadre duquel nous collaborons aussi avec l'équipe "Les Décodeurs" (journalistes spécialisés dans le fact-checking) du journal Le Monde.La démarche scientifique de la thèse se décompose comme suit:1. Identifier les technologies et domaines de gestion de contenu (texte, données, connaissances) intervenant de façon recurrente (ou dont le besoin est ressenti comme important) dans l'activité des journalistes.Il est par exemple déjà clair que ceux-ci ont l'habitude d'utiliser "en interne" quelques bases de données construites par les journalistes eux-mêmes ; ils disposent aussi d'outils internes (à la rédaction) de recherche par mots-clé ; cependant, ils souhaiterait augmenter leur capacité d'indexation sémantique...Parmi ces problèmes, identifier ceux pour lesquels des solutions techniques (informatiques) sont connues, et le cas échéant mis en oeuvre dans des systèmes existants.2. S'attaquer aux problèmes ouverts (sur le plan de la recherche), pour lesquels des réponses satisfaisantes manquent, liés à la modélisation et à l'algorithmique efficace pour des contenus textuels, sémantiques, et des données, dans un contexte journalistique. / Digital content is increasingly produced nowadays in a variety of media such as news and social network sites, personal Web sites, blogs etc. In particular, a large and dynamic part of such content is related to media-worthy events, whether of general interest (e.g., the war in Syria) or of specialized interest to a sub-community of users (e.g., sport events or genetically modified organisms). While such content is primarily meant for the human users (readers), interest is growing in its automatic analysis, understanding and exploitation. Within the ANR project ContentCheck, we are interested in developing textual and semantic tools for analyzing content shared through digital media. The proposed PhD project takes place within this contract, and will be developed based on the interactions with our partner from Le Monde. The PhD project aims at developing algorithms and tools for :Classifying and annotating mixed content (from articles, structured databases, social media etc.) based on an existing set of topics (or ontology) ;Information and relation extraction from a text which may comprise a statement to be fact-checked, with a particular focus on capturing the time dimension ; a sample statement is for instance « VAT on iron in France was the highest in Europe in 2015 ».Building structured queries from extracted information and relations, to be evaluated against reference databases used as trusted information against which facts can be checked.

<div><div><div><p>Electronic health records collect an enormous amount of data about patients. However, the information about the patient’s illness is stored in progress notes that are in an un- structured format. It is difficult for humans to annotate symptoms listed in the free text. Recently, researchers have explored the advancements of deep learning can be applied to pro- cess biomedical data. The information in the text can be extracted with the help of natural language processing. The research presented in this thesis aims at automating the process of symptom extraction. The proposed methods use pre-trained word embeddings such as BioWord2Vec, BERT, and BioBERT to generate vectors of the words based on semantics and syntactic structure of sentences. BioWord2Vec embeddings are fed into a BiLSTM neural network with a CRF layer to capture the dependencies between the co-related terms in the sentence. The pre-trained BERT and BioBERT embeddings are fed into the BERT model with a CRF layer to analyze the output tags of neighboring tokens. The research shows that with the help of the CRF layer in neural network models, longer phrases of symptoms can be extracted from the text. The proposed models are compared with the UMLS Metamap tool that uses various sources to categorize the terms in the text to different semantic types and Stanford CoreNLP, a dependency parser, that analyses syntactic relations in the sentence to extract information. The performance of the models is analyzed by using strict, relaxed, and n-gram evaluation schemes. The results show BioBERT with a CRF layer can extract the majority of the human-labeled symptoms. Furthermore, the model is used to extract symptoms from COVID-19 tweets. The model was able to extract symptoms listed by CDC as well as new symptoms.</p></div></div></div>

[ES] En esta tesis abordamos el problema de la detección de las opiniones (stance detection, SD) en las redes sociales, centrándose en los debates políticos polarizados en Twitter. La SD consiste en determinar automáticamente si el autor de una publicación está a favor o en contra de un objetivo de interés, o si no se puede inferir la opinión. Nos ocupamos de temas políticos como las elecciones políticas y los referendos y, como resultado, los objetivos son tanto personas como referendos. También exploramos las comunicaciones que tienen lugar en estos debates polarizados, arrojando luz sobre las dinámicas de comunicación entre personas que tienen opiniones en acuerdo o en conflicto, enfocándonos en particular en la observación del cambio de opiniones (opinion shifting). Proponemos modelos de aprendizaje automático para la SD como si fuera un problema de clasificación binaria. Exploramos características basadas en el contenido del texto del tweet, además usamos características basadas en información contextual que no emerge directamente del texto. Utilizando el corpus de benchmark propuesto para la tarea compartida sobre la SD realizado para SemEval 2016, exploramos la contribución que el estudio de las relaciones entre el objetivo de interés y las otras entidades involucradas en el debate proporciona a la SD. Al participar en la tarea ``Stance and Gender Detection in Tweets on Catalan Independence'' organizado para IberEval 2017, hemos propuesto otras características textuales y contextuales para la SD en tweets en español y en catalán. Explorando la SD desde una perspectiva multilingüe, hemos creado un corpus de tweets en francés y uno en italiano. La extensión multilingüe de nuestro modelo (multiTACOS) muestra que la SD está influenciada más por los diferentes estilos utilizados por los usuarios para comunicar la opinión sobre objetivos de diferentes tipos (personas o referendos) en lugar del idioma utilizado. Con el objetivo de recuperar información contextual sobre la red social de los usuarios de Twitter (generalmente las tareas compartidas solo consisten en el contenido del tweet, dejando de lado la información sobre el usuario), hemos creado otros dos conjuntos de datos, uno en inglés y uno en italiano, respectivamente, sobre el Brexit (TW-BREXIT) y sobre el referéndum constitucional italiano (ConRef-STANCE-ita). En ambos casos de estudio, mostramos que los usuarios tienden a agruparse en grupos con ideas similares. Por este motivo, el modelo que explota el conocimiento de la comunidad social a la que el autor del tweet pertenece, supera los resultados obtenidos utilizando solo las funciones basadas en el contenido de la publicación. Además, la evidencia muestra que los usuarios utilizan diferentes tipos de comunicación según el nivel de acuerdo con la opinión del interlocutor, por ejemplo, las relaciones de amistad, los retweets y las citas (quote) son más comunes entre los usuarios relacionados, mientras que las respuestas (replies) se utilizan a menudo para interactuar con usuarios que tienen diferentes posiciones. Al abordar la SD desde una perspectiva diacrónica, también observamos tanto el cambio de opinión como la mitigación del debate hacia posiciones neutrales después del resultado de la votación. Además, hemos observado que tener contacto con una variedad más amplia de opiniones puede influir en la propensión a cambiar de opinión. Finalmente, mostramos que las características basadas en una representación gráfica de un dominio de interés no se limitan a la SD, sino que se puede aplicar a diferentes escenarios. Al proponer otra tarea de clasificación que realiza la identificación del talento en el deporte, especialmente en el estudio de caso del tenis de mesa, mostramos que las métricas de redes basadas en la centralidad son una señal fuerte para el talento y pueden usarse para entrenar un modelo de algoritmo de aprendizaje automático para enfrentar esta / [CAT] En aquesta tesi doctoral abordem el problema de la detecció de posició (stance detection, SD) en els mitjans de comunicació social, especialment centrat en els debats polítics polaritzats a Twitter. La SD consisteix a determinar automàticament si l'autor d'una publicació està a favor o en contra d'un objectiu o tema d'interès, o si l'opinió envers d'aquest objectiu o tema determinat no es pot inferir. Ens ocupem de temes polítics com ara esdeveniments electorals i, en conseqüència, els temes d'interès són, en concret, la SD en vers dirigents polítics i referèndums. També explorem les comunicacions que es duen a terme en aquests debats polaritzats, que posen de manifest la dinàmica de les comunicacions entre les persones que tenen opinions concordants o contrastades, especialment centrant-nos en l'observació del canvi de les opinions. Proposem models d'aprenentatge automàtic per abordar la SD com un problema de classificació. Explorem les funcions basades en el contingut textual del tweet, però també les funcions basades en la informació contextual que no afloren directament del text. Utilitzem el conjunt de dades de referència en anglès proposat per a les tasques compartides sobre SD celebrades a SemEval 2016, per explorar la contribució a la SD d'investigar les relacions entre l'objectiu d'interès i les altres entitats implicades en el debat. En la participació a la tasca compartida de ``Stance and Gender Detection in Tweets on Catalan Independence'' celebrada a IberEval 2017, es van proposar altres trets textuals i contextuals per detectar la posició dels autors dels tweets, escrits en espanyol i en català, envers la independència de Catalunya. L'extensió multilingüe del model de SD (multiTACOS) mostra que la SD es veu afectada pels diferents estils que utilitzen els usuaris per comunicar la posició envers objectius de diferents tipus (persones o referèndum) més que la llengua utilitzada. Amb l'objectiu de recuperar informació contextual sobre la xarxa social dels usuaris de Twitter (les tasques compartides solen publicar només el contingut del tweet i deixen de banda, en canvi, la informació sobre la persona que escriu el tweet), vam crear dos conjunts més de dades, un en anglès i un en italià, el corpus Brexit (TW-BREXIT) i el corpus del referèndum constitucional italià (ConRef-STANCE-ita) respectivament. En els dos casos, demostrem que els usuaris tendeixen a agrupar-se en grups d'opinió o creences similars. Per aquest motiu, el model aprofita el coneixement de la comunitat social en línia al qual pertany el tweeter i supera els resultats obtinguts utilitzant només funcions basades en el contingut de la publicació. És més, els experiments també mostren que els usuaris fan servir diferents tipus de comunicació en funció del nivell d'acord amb l'opinió del seu interlocutor, és a dir, les relacions d'amistat (friendship), retweets i cotitzacions (quotes) són més freqüents entre els usuaris amb idees afins, mentre que les respostes (replies) s'utilitzen sovint per interactuar amb els usuaris que tenen posicions o opinions diferents. A l'hora d'abordar la SD des d'una perspectiva diacrònica, també observem el canvi d'opinió i la mitigació del debat cap a una posició no alineament després del resultat de la votació. A continuació, observem que l'accés a una major diversitat de punts de vista pot influir en la propensió a canviar l'opinió personal. Finalment, mostrem que la utilitat de les funcions basades en una representació gràfica d'un domini d'interès no es limita a la SD, sinó que es pot aplicar a diferents escenaris. Proposar una altra tasca de classificació que realitzi la identificació de talent en l'esport, especialment centrada en l'estudi de cas del tennis de taula, mostrem que les xarxes mètriques basades en la centralitat són un fort senyal per a detectar el talent i també es pot utilitzar per a l'entrenament d'un model d'algorisme d'ap / [EN] In this thesis, we address the problem of stance detection (SD) in social media focusing on polarized political debates in Twitter. SD consists in automatically determine whether the author of a post is in favor or against a target of interest, or whether the opinion toward the given target can not be inferred. We deal with political topics such as electoral events and consequently the targets of interest are both politicians and referendums. We also explore the communications which take place in these polarized debates shedding some light on dynamics of communications among people having concordant or contrasting opinions, particularly focusing on observing opinions' shifting. We propose machine learning models for addressing SD as a classification problem. We explore features based on the textual content of the tweet, but also features based on contextual information that do no emerge directly from the text. Using the English benchmark dataset proposed for the shared tasks on SD held at SemEval 2016, we explore the contribution on SD of investigating the relations among the target of interest and the other entities involved in the debate. Participating to the ``Stance and Gender Detection in Tweets on Catalan Independence'' shared task held at IberEval 2017, we proposed other textual and contextual based features for detecting stance on Spanish and Catalan tweets. With the main aim of facing SD in a multilingual perspective and having an homogeneous setting for multi-language comparisons, we collected tweets in French and Italian also. The multilingual extension of our SD model (multiTACOS) shows that SD is affected by the different styles used by users for communicating stance towards target of different types (persons or referendum) more than the used language. With the aim of retrieving contextual information about the social network of Twitter's users, we created other two datasets, one in English and one in Italian, respectively about the Brexit (TW-BREXIT) and the Italian Constitutional referendum (ConRef-STANCE-ita). In both the case studies, we show that users tend to aggregate themselves in like-minded groups. For this reason, the model takes advantage of knowing the online social community the tweeter belongs to and outperforms the results obtained by using only features based on the content of the post. Furthermore, experiments show that users use different type of communication depending on the level of agreement with the interlocutor's opinion, i.e., friendship, retweets, and quote relations are more common among like-minded users, while replies are often used for interacting with users having different stances. Addressing SD in a diachronic perspective, we also observe both opinion shifting and a mitigation of the debate towards an unaligned position after the outcome of the vote. Then, we observe that accessing to a larger diversity of point of views can influence the propensity to change the personal opinion. We finally show that the usefulness of features based on a graph representation of a domain of interest is not limited to SD, but can be applied to different scenarios. Proposing another classification task that performs talent identification in sport, particularly focusing on the case study of table tennis, we show that networks metrics based on centrality are strong signal for talent and can be used for training a machine learning algorithm model for this task too. / Lai, M. (2019). On Language and Structure in Polarized Communities [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/119116 / TESIS

Pharmacovigilance, also referred to as drug safety, is an important science for identifying risks related to medicine intake. Side effects of medicine can be caused by for example interactions, high dosage and misuse. In order to find patterns in what causes the unwanted effects, information needs to be gathered and mapped to predefined terms. This mapping is today done manually by experts which can be a very difficult and time consuming task. In this thesis the aim is to automate the process of mapping side effects by using machine learning techniques. The model was developed using information from preexisting mappings of verbatim expressions of side effects. The final model that was constructed made use of the pre-trained language model BERT, which has received state-of-the-art results within the NLP field. When evaluating on the test set the final model performed an accuracy of 80.21%. It was found that some verbatims were very difficult for our model to classify mainly because of ambiguity or lack of information contained in the verbatim. As it is very important for the mappings to be done correctly, a threshold was introduced which left for manual mapping the verbatims that were most difficult to classify. This process could however still be improved as suggested terms were generated from the model, which could be used as support for the specialist responsible for the manual mapping.

Modern deep learning methods have boosted research progress in visual recognition and text understanding but it is a non-trivial task to unite these advances from both disciplines. In this thesis, we develop models and techniques that allow us to connect natural language and visual content enabling automatic video subtitling, visual grounding, and text-based image search. Such models could be useful in a wide range of applications in robotics and human-computer interaction bridging the gap in vision and language understanding. First, we develop a model that generates natural language descriptions of the main activities and scenes depicted in short videos. While previous methods were constrained to a predefined list of objects, actions, or attributes, our model learns to generate descriptions directly from raw pixels. The model exploits available audio information and the video’s category (e.g., cooking, movie, education) to generate more relevant and coherent sentences. Then, we introduce a technique for visual grounding of generated sentences using the same video description model. Our approach allows for explaining the model’s prediction by localizing salient video regions for corresponding words in the generated sentence. Lastly, we address the problem of image retrieval. Existing cross-modal retrieval methods work by learning a common embedding space for different modalities using parallel data such as images and their accompanying descriptions. Instead, we focus on the case when images are connected by relative annotations: given the context set as an image and its metadata, the user can specify desired semantic changes using natural language instructions. The model needs to capture distinctive visual differences between image pairs as described by the user. Our approach enables interactive image search such that the natural language feedback significantly improves the efficacy of image retrieval. We show that the proposed methods advance the state-of-the-art for video captioning and image retrieval tasks in terms of both accuracy and interpretability.

Native Language Identification (NLI) is the task of predicting the native language of an author from their text written in a second language. The idea is to find writing habits that transfer from an author’s native language to their second language. Many approaches to this task have been studied, from simple word frequency analysis, to analyzing grammatical and spelling mistakes to find patterns and traits that are common between different authors of the same native language. This can be a very complex task, depending on the native language and the proficiency of the author’s second language. The most common approach that has seen very good results is based on the usage of n-gram features of words and characters. In this thesis, we attempt to extract lexical, grammatical, and semantic features from the sentences of non-native English essays using neural networks. The training and testing data was obtained from a large corpus of publicly available essays written by authors of several countries around the world. The neural network models consisted of Long Short-Term Memory and Convolutional networks using the sentences of each document as the input. Additional statistical features were generated from the text to complement the predictions of the neural networks, which were then used as feature inputs to a Support Vector Machine, making the final prediction. Results show that Long Short-Term Memory neural network can improve performance over a naive bag of words approach, but with a much smaller feature set. With more fine-tuning of neural network hyperparameters, these results will likely improve significantly.