Brazil

Neural machine translation (NMT) nowadays requires an increasing
amount of data and computational power, so succeeding in this task with limited
data and using a single GPU might be challenging. Strategies such as the use
of pre-trained word embeddings, subword embeddings, and data augmentation
solutions can potentially address some issues faced in low-resource experimental settings, but their impact on the quality of translations is unclear. This work
evaluates some of these strategies on two low-resource experiments beyond just
reporting BLEU: errors are categorized on the Portuguese-English pair with the
help of a translator, considering semantic and syntactic aspects. The BPE subword approach has shown to be the most effective solution, allowing a BLEU
increase of 59% p.p. compared to the standard Transformer.

BRACIS 2021

Tackling neural machine translation in low-resource settings: a Portuguese case study

natural language processing

neural machine translation

deep learning

**BRACIS** originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence – **SBIA** (22 editions), and the Brazilian Symposium on Neural Networks – **SBRN** (13 editions). **ENIAC** is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. **STIL** is the bi-annual Language Technology event.
All events are supported by the Brazilian Computer Society – Special Interest Groups: **CE-IA**, **CE-IC** and **CE-PLN**.

To access the 10th Brazilian Conference on Intelligent Systems (**BRACIS**), you need to register for the conference.  Please visit the Conference official site to register

[Register now](http://c4ai.inova.usp.br/bracis2021/)

Register now

BRACIS originated from the combination of the two most important scientific events in Brazil in Artificial Intelligence (AI) and Computational Intelligence (CI): the Brazilian Symposium on Artificial Intelligence – SBIA (22 editions), and the Brazilian Symposium on Neural Networks – SBRN (13 editions). 
ENIAC is a national forum for researchers, practitioners, educators and students to present and discuss innovations, trends, experiences, developments and work in progress in the fields of Artificial and Computational Intelligence. 
STIL is the bi-annual Language Technology event.
All events are supported by the Brazilian Computer Society – Special Interest Groups: CE-IA, CE-IC and CE-PLN

technical paper

In this research, qualitative and quantitative characteristics of literary genres of the BDCamões corpus are evaluated under the bottom-up perspective. We show that textual genres make information more easily understood by search tools. We started from chronicles, novels, short stories, and tales annotated according to the technique of universal dependence classified according to random decision forests. The use of the LIWC dictionary in Brazilian-Portuguese gave access to psycholinguistic resources. The results are presented in two tables: one with the results by class reported by the mean together with a measure of variability; and another with the importance of characteristics by class, LIWC labels, part of speech and universal dependency labels, highlighting the ten highest positive characteristics and the ten lowest negative characteristics. These tools have adapted to the contextual characteristics of fluidity and changeability of literary genres, overcoming the difficulty of standard NLP varieties These results have potential uses to replicate contextual characteristics, contemplating the artistic content of texts, which makes the machine more intuitive and with more consistent results.

A machine learning approach to literary genre classification on Portuguese texts: circumventing NLP's standard varieties

Automatic text summarization aims at condensing the contents of a text into a simple and descriptive summary. Summarization techniques drastically benefited from the recent advances in Deep Learning. Nevertheless, these techniques are still unable to properly deal with long texts. In this work, we investigate whether the combination of summaries extracted from multiple sections of long scientific texts may enhance the quality of the summary for the whole document. We conduct experiments on real world corpus to assess the effectiveness of our proposal. The results show that our multi-section proposal are as good as summaries generated using the entire text as input and twice as good as single section.

A Long Texts Summarization Approach to Scientific Articles

In recent years, researchers in the area of Computational Creativity have studied the human creative process proposing different approaches to reproduce it with a formal procedure. We introduce a model for the generation of literary rhymes in Spanish, combining structures of language and neural network models Word2vec, into a structure for semantic assimilation. The results obtained with a manual evaluation of the texts generated by our algorithm are encouraging

A Preliminary Study for Literary Rhyme Generation based on Neuronal Representation, Semantics Resources and Shallow Parsing

The production of fake news is a problem nowadays. With social networks, fake news spread in easier and less costly ways, having the power to reach a large number of people in a short time. In this paper, we investigate graph-based approaches for fake news characterization and detection, taking into account widely used measures of graphs and complex networks. Our results show that some network measures are useful for structurally characterizing fake and true news and that machine learning-based solutions over this kind of feature produce promising results.

Structural Characterization and Graph-based Detection of Fake News in Portuguese

Professional fact-checking tends to be costly and scalability challenges. For this reason, a series of methods to automate this process has been emerging. However, these methods have been created with monolithic architectures, not making use of pre-built parts, and are harder to be understood by others. This work aims to propose a framework, where the development of methods can be done in a modular manner, creating workflows based on key steps that can be linked together to generate the input classification. The framework makes use of a proposed data traceability ontology to map all generated data during execution under a unified vocabulary, which facilitates communication between these independent components.

ReVera Framework: Um Framwork para rastreabilidade em fact-checking automático

In this presentation we show an empirical study of document retrieval and reader for automatically answering conceptual student questions about technology and management courses in an online education platform.

An Empirical Study of Information Retrieval and Machine Reading Comprehension Algorithms for an Online Education Platform

The main purpose of stemming is to reduce the inflected words into its root form or stem. Thus, words can be mapped to the same concept, improving the process of information retrieval, regarding its ability to index documents and to reduce data dimensionality. However, the efficiency of those algorithms varies according to different aspects. Also, studies in the field area reached contrasting conclusions. This work assesses the use of stemmers in the retrieval of legislative documents written in Portuguese. Four stemmers together with BM25 were evaluated in two legislative corpora from the Brazilian Chamber of Deputies. RSLP-S and Savoy stemmers showed the best improvements in the information retrieval pipeline.

Assessing the Impact of Stemming Algorithms Applied to Brazilian Legislative Documents Retrieval

In this work, we carried out a study about the use of attention-based algorithms to automate the categorization of Brazilian case law documents. We used data from the Kollemata Project to produce two distinct datasets with adequate class systems. Then, we implemented a multi-class and multi-label version of BERT and fine-tuned different BERT models with the produced datasets. We evaluated several metrics, adopting the micro-averaged F1-Score as our main metric for witch we obtained a performance value of F1=0.72, corresponding to significative gains over the tested statistical baseline.

verBERT: Automating Brazilian Case Law Document Multi-label Categorization Using BERT

State-of-the-art models have obtained high accuracy on mainstream Natural Language Inference (NLI) datasets. However, recent research has suggested that the task is far from solved. Current models struggle to generalize and fail to consider the inherent human disagreements in tasks such as NLI. In this work, we conduct an experiment based on a small subset of the NLI corpora such as SNLI and SICK. It reveals that some inference cases are inherently harder to annotate than others, although good-quality guidelines can reduce this difficulty to some extent. We propose adding a Difficulty Score to NLI datasets, to capture the human difficulty level of agreement.

Annotation Difficulties in Natural Language Inference

A chatbot is an artificial intelligence based system aimed at chatting with users, commonly used as a virtual assistant to help people or answer questions. Intent classification is an essential task for chatbots where it aims to identify what the user wants in a certain dialogue. However, for many domains, little data are available to properly train those systems. In this work, we evaluate the performance of two methods to generate synthetic data for chatbots, one based on template questions and another based on neural text generation. We build four datasets that are used training chatbot components in the intent classification task. We intend to simulate the task of migrating a search-based portal to an interactive dialogue-based information service by using artificial datasets for initial model training. Our results show that template-based datasets are slightly superior to those neural-based generated in our application domain, however, neural-generated present good results and they are a viable option when one has limited access to domain experts to hand-code text templates.

Premium content

Tackling neural machine translation in low-resource settings: a Portuguese case study

Downloads

Next from BRACIS 2021

A machine learning approach to literary genre classification on Portuguese texts: circumventing NLP's standard varieties

Similar lecture

A Deep Learning approach for Aspect Sentiment Triplet Extraction in Portuguese

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES