The objective of this work is to apply together ontology with bag-of-words models, similarity learning, and document classification in texts with uttered decisions. The objective is to improve the results of data mining in a database of court decisions. An automatic method of searching sentences in judicial processes related to the one under judgment is developed using the frequency term-inverse of frequency in documents model together with the Jaccard similarity coefficient, establishing weights on the co-occurrence of terms in legal texts of the same category. A dataset with document vectorization is used for supervised training of machine learning algorithms, aiming to classify new justice processes. The proposed methodology provides flexibility to the Judiciary, simulating the role of legal advisors in preparing court decisions with less time and efficiency in the search for jurisprudential standards. The results obtained show that, through accuracy metrics, the proposed model is effective and efficient, and can be applied in the process of identification of court decisions. Thus, the application of artificial intelligence, ontology, and data mining is indicated for information retrieval in court decisions.

We witness an overload of textual data: a vast amount of information is recorded in natural language text and stored in digital media. In the life sciences fields, the increasing number of biomedical scientific publications and of clinical reports retains a wealth of knowledge that must be unearthed and linked through automatic information extraction methods. These are imperative to assist curation in biological databases and play an important role in drug discovery, precision medicine, and pharmacological and clinical research. This thesis investigates the use of natural language processing, machine learning, and knowledge-based methods to extract information from biomedical text in English language. Specifically, we study and propose methods for entity disambiguation, document classification, and relation extraction. Overall, this work contributes with an exhaustive evaluation study of several approaches for distinct biomedical information extraction tasks, which are a vital support for the advancement of the current knowledge.

Many text classification problems in social networks, and other contexts, are also dynamic problems, where concepts drift through time, and meaningful labels are dynamic. In Twitter-based applications in particular, ensembles are often applied to problems that fit this description, for example sentiment analysis or adapting to drifting circumstances. While it can be straightforward to request different classifiers' input on such ensembles, our goal is to boost dynamic ensembles by combining performance metrics as efficiently as possible. We present a twofold performance-based framework to classify incoming tweets based on recent tweets. On the one hand, individual ensemble classifiers' performance is paramount in defining their contribution to the ensemble. On the other hand, examples are actively selected based on their ability to effectively contribute to the performance in classifying drifting concepts. The main step of the algorithm uses different performance metrics to determine both each classifier strength in the ensemble and each example importance, and hence lifetime, in the learning process. We demonstrate, on a drifted benchmark dataset, that our framework drives the classification performance considerably up for it to make a difference in a variety of applications.

Drift is a given in most machine learning applications. The idea that models must accommodate for changes, and thus be dynamic, is ubiquitous. Current challenges include temporal data streams, drift and non-stationary scenarios, often with text data, whether in social networks or in business systems. There are multiple drift patterns types: concepts that appear and disappear suddenly, recurrently, or even gradually or incrementally. Researchers strive to propose and test algorithms and techniques to deal with drift in text classification, but it is difficult to find adequate benchmarks in such dynamic environments. In this paper we present DOTS, Drift Oriented Tool System, a framework that allows for the definition and generation of text-based datasets where drift characteristics can be thoroughly defined, implemented and tested. The usefulness of DOTS is presented using a Twitter stream case study. DOTS is used to define datasets and test the effectiveness of using different document representation in a Twitter scenario. Results show the potential of DOTS in machine learning research.

Atualmente, os Sistemas de Recomendação têm apresentado um aumento na sua utilização, uma vez que permitem uma melhor entrega de um serviço aos seus utilizadores. No entanto, de forma a que estes sistemas consigam ser robustos e eficientes, é necessária a existência de uma base de conteúdo corretamente classificado. No caso de classificação de programas de televisão, a sua classificação pode, para além de melhorar o conteúdo mostrado no Guia de Programas Eletrónicos, melhorar a base de dados utilizada para os Sistemas de Recomendação e consequentemente o sistema que faz uso desses dados. De forma a automatizar o processo, o Processamento de Linguagem Natural surge como um conjunto de técnicas muito útil, uma vez que permite às máquinas entender e processar linguagem natural, assim como classificar grandes quantidades de dados automaticamente. Este documento, elaborado em contexto de Dissertação para obtenção de Mestrado e estágio na empresa Altice Labs em articulação com a Altice/MEO, apresenta o processo de desenvolvimento de um modelo capaz de classificar programas de televisão pela sua categoria, com base no texto da sua sinopse, utilizando técnicas de Processamento de Linguagem Natural. Começamos por discutir o contexto teórico de alguns conceitos importantes para o problema, seguindo-se uma revisão a trabalhos semelhantes de forma aperceber as técnicas tipicamente implementadas e resultados obtidos. Seguidamente, são apresentadas, com detalhe, as tarefas desenvolvidas e decisões tomadas no contexto desta investigação/estágio, às quais acresce uma explicação de como o produto final está a ser integrado nos sistemas da MEO para ser utilizado diretamente nas suas bases de dados. O desenvolvimento de um modelo inteligente para classificação de programas permite à Altice/MEO a automatização dos seus processos de classificação e um consequente melhoramento do conteúdo apresentado no Guia de Programas Eletrónicos (EPG) e nos Sistemas de Recomendação que fazem uso dessa classificação, assim como ajudará também a indicar oportunidades de negócio ao melhorar o detalhe no perfil dos clientes.

A Análise de Sentimento em diálogo visa detectar o sentimento expresso em frases e, da forma em que foi aplicada neste trabalho, tem o potencial de melhorar a interação humano-computador em linguagem natural. Nesta dissertação, exploramos diferentes abordagens para Análise de Sentimento em diálogos escritos em Português, principalmente no domínio das Telecomunicações. Se integrado num agente conversacional, isto permitirá a identificação automática e uma reação rápida a clientes que manifestam sentimento negativo, possivelmente com intervenção humana, podendo assim minimizar os danos. Foram criados dois datasets de dados reais, com sentimento manualmente anotado: um com diálogos de call center, cedido pela Altice Labs (AL); outro a partir de conversas extraídas do Twitter, envolvendo principalmente contas de empresas de Telecomunicações. Comparamos o desempenho de diferentes abordagens com complexidades variáveis, desde modelos baseados em léxico, passando por classificadores mais tradicionais (e.g., Random Forest, Regressão Logística), até modelos mais recentes (e.g., Fine-tuned Bidirectional EncoderRepresentations from Transformers (BERT), Few-Shot Learning). Como um diálogo é uma sequência de falas, as frases anteriores podem ter impacto no sentimento da frase atual. Assim sendo, também desenvolvemos modelos que consideram contexto (e.g., BERT-Confitional Random Field (BERT-CRF)). Cada modelo de aprendizagem computacional, com a excepção do último grupo, é analisado com e sem a inclusão de falas anteriores. Ao classificar os datasets, o melhor modelo (Fine-tuned BERT) atingiu F1-Scores de 0,87 e 0,93 nos datasets da Altice Labs e do Twitter, respetivamente. O desempenho do primeiro foi alcançado sem considerar contexto, enquanto o do segundo foi alcançado considerando contexto (inclusão das falas anteriores). No entanto, na maioria dos cenários, a utilização de contexto parece diminuir o desempenho dos classificadores, significando que, para esta aplicação, a utilização da fala atual pode ser suficiente. Estes são resultados interessantes que sugerem que um apoio ao cliente automático pode beneficiar de um componente de análise de sentimento. A abordagem desenvolvida será disponibilizada à AL, que poderá considerar a sua integração no seu sistema de assistência ao cliente.

Social media is a great source of data for analyses, since they provide ways for people to share emotions, feelings, ideas, and even symptoms of diseases. By the end of 2019, a global pandemic alert was raised, relative to a virus that had a high contamination rate and could cause respiratory complications. To help identify those who may have the symptoms of this disease or to detect who is already infected, this paper analyzed the performance of eight machine learning algorithms (KNN, Naive Bayes, Decision Tree, Random Forest, SVM, simple Multilayer Perceptron, Convolutional Neural Networks and BERT) in the search and classification of tweets that mention self-report of COVID-19 symptoms. The dataset was labeled using a set of disease symptom keywords provided by the World Health Organization. The tests showed that Random Forest algorithm had the best results, closely followed by BERT and Convolution Neural Network, although traditional machine learning algorithms also have can also provide good results. This work could also aid in the selection of algorithms in the identification of diseases symptoms in social media content.

The use of dialogue agents for common tasks is increasing. The use of chatbots is also popular, an indication of humans' willingness to partake in a conversation with a “robot". Despite their appealingness, interactions with conversational agents are limited and many times cause negative reactions from humans. Sentiment analysis is a process that tries to find expressions of human emotions and feelings or general sentiments from sources such as speech, text, or body language. The objective for the work presented is to provide tools to make it possible to integrate information regarding the emotional state of the user of a conversational agent to improve interaction, reduce negative reactions and improve user experience. This work, in collaboration with GoContact, a company that delivers a cloud-based Contact Center as a service solution, with automation powered by AI, contemplated the development of an initial version of a sentiment analysis module aligned with requirements derived from scenarios based on GoContact experience. This first iteration processes the speech-to-text word sequences and is based on LinguaKit and MeaningCloud. The work also led to the development of a rudimentary bot, made to recreate the conditions described in the scenarios and to gather a dataset, essential for module evaluation and future evolution of the module (e.g. integration of speech analysis). The module was evaluated with the acquired dataset, providing useful information regarding the potential of LinguaKit and MeaningCloud for the sentiment analysis of interactions with conversational agents in Portuguese.