Trabalhos Aceitos

1. A Brief Survey of Deep Learning based methods, against OpenNLP NameFinder for Named Entity Recognition on Portuguese Literary Texts

• Autores: Vinicius Amaro Sampaio, Mardônio J. C. França, Paulo Bruno Lopes da Silva, Gustavo Augusto Lima de Campos, Lara Domingos Hissa
• Abstract: This paper is a brief comparative study of Named Entity Recognition (NER). It presents a evaluation of Apache's OpenNLP against CRF and Deep Learning methods and frameworks such as Tensorflow. This evaluation is done by looking at the F1-score. For this work we tagged and train our own models using samples of literary texts from Machado de Assis e Jayson Aguiar. We then see the potential of Deep Learning methods and plan on investigating more recent approaches and frameworks like Pytorch and LSTM+CRF, also we plan to expand the size of our corpus.

3. Chatbot para auxiliar os discentes nos procedimentos administrativos de uma universidade

• Autores: Wesley Benício dos Santos Silva, Márcio de Souza Dias, Nádia F. F. da Silva
• Title: Chatbot to help students in the administrative procedures of a university
• Abstract: Most new students of a university are unaware functioning of the university, which can cause inconvenience in requesting a service offered by the institution. Thus, the purpose of this project will be to develop a chatbot tool that will help students with questions about university procedures. This tool will seek to respond quickly and friendly to questions from students regarding the various procedures involved in the university, taking the necessary steps to resolve these questions.

5. Compilação de um Banco Multilíngue de Acolhimento a Pessoas Refugiadas

• Autores: Anna B. D. Furtado, Elisa D. Teixeira
• Title: Compilation of a Multilingual Databank for Sheltering Refugees
• Abstract: The migration crisis can be seen in every continent of the world. In Brazil, it is not different. In 2016, the country received 10.308 asylum claims. In this context, the MOBILANG (Mobilities and Languages in Contact) research group proposes to create a multilingual terminological databank from frequent word clusters obtained through texts and material refugees receive when they arrive in Brazil. In this work, we attempt to systematize the steps taken to compile a multilingual corpus on migration and asylum and extract clusters using Corpus Linguistics as our theoretical and methodological approach. We extract term candidates with the Sketch Engine software. As results, we present data on the corpus and a small glossary using the term candidates and their respective translations. Finally, we discuss the results reliability and future steps.

7. ET: uma Estação de Trabalho para revisão, edição e avaliação de corpora anotados morfossintaticamente

• Autores: Elvis de Souza, Cláudia Freitas
• Title: A workstation for revising, editing and evaluating morphosyntactically annotated corpora
• Abstract: Morphosyntactic annotation systems that use machine learning technology require large and well-annotated corpora to learn how to annotate texts properly. In general, to improve quality of automatic annotators, much has been done in relation to the technology behind the systems, however, there are still some bottlenecks in the quality of the material that serves as training, which ends up impacting negatively the quality of learning. We believe that a way to overcome these bottlenecks may be linguistically: improving the annotation of training corpora, making them more consistent and eliminating possible human errors. In this context, we present here a workstation drawn from the linguistic perspective with the objective of facilitating the revision, the editing and the evaluation of annotated corpora, aligning the work done by the language experts, on the one hand, and the practical results, that is, the performance of NLP systems, on the other hand. This way, theoretical discussions about grammatical categories may be based not only on language suitability to certain theories, but also on the empirical results of machine learning systems.

9. Investigação do uso de word embeddings para cálculo de similaridade em memórias de tradução

• Autores: Karina Mayumi Johansson, Helena de Medeiros Caseli
• Title: Research of the use of word embeddings for calculation of similarity in translation memories
• Abstract: The strategy traditionally employed by the CAT tools to match the segments of the phrase being currently translated with the segments present in the translation memory considers the intersection of the sequence of words (n-grams) present in the segments of the text being compared. However, this strategy is not capable of capturing semantic similarities beyond the trivial level. This study therefore presents a project with the aim of investigating the applicability of monolingual and bilingual word embeddings to implement the matching. The study is still in its initial phase of development. In sequence, there will be proposed and implemented a strategy for the calculation of similarity using word embeddings, which will be incorporated in a open source CAT tool. In order to evaluate the proposed strategies, the quality of matching in the baseline system (a version of a CAT system without any modification) will be compared to those of the system in which the proposed method will be implemented. At the conclusion of this project is expected to have obtained a strategy based on semantic similarity that will be an alternative to the traditional matching strategy based on n-grams. Although there are already texts covering the use of word embeddings to detect the textual similarity and cleaning of translation memories, there is no literature about any work that has investigated the objective of this project. Consequently, this study should be considered as the first initiative to an investigation within this context.

11. Preparação para Leitura Distante em português: diálogos entre PLN e Humanidades Digitais

• Autores: Luísa Rocha, Cláudia Freitas, Diana Santos
• Title: Preparing for Distant Reading in Portuguese: the dialogue between NLP and Digital Humanities
• Abstract: Distant Reading is a technique developed by Franco Moretti that means to analyze patterns in texts of the literary genre. To apply this technique to literary works of the Portuguese language, we found it was important to survey a set of works on the matter, thus becoming more familiar with the area. Consequently, we've noticed that a few adjustments had to be made, and precautions had to be taken, when handling the OBras corpus. The correction of the grammatical genre of proper nouns, the addition of a semantic role, and the correction of the segmentation were the adjustments made and addressed in this paper so that exploring the OBras corpus can happen in the best way possible.

13. Um estudo sobre desidentificação de evoluções clínicas

• Autores: Thaila Elisa Quaini, Henrique D. P. dos Santos, Sandra C. de Abreu, Bernardo S. Consoli and Renata Vieira
• Title: A study on de-identification of clinical notes
• Abstract: Medical records of patients are important in the field of medical research. However, to obtain a patient identity, a Health Insurance Portability and Accountability Act (HIPAA) is required, which must be removed prior to the survey. Manual de-identification on large amounts of medical records data is expensive, time-consuming and error prone, requiring large-scale automated de-identification methods. This paper presents an analysis of the problem of the Brazilian Portuguese language, for a task of disidentification of electronic medical records. We compare the main types of business rule identification with an approach based on a list of names specially built for a task. The list of names was developed from the database and using the word Embedded to specialize the names through the semantic similarity between words.