Brazil

The iris flower dataset is a ubiquitous benchmark task in machine learning literature. With its 150 instances, four continuous features, and three balanced classes, of which one is linearly separable from the others, the iris is generally considered an easy problem. Hence researchers usually rely on other datasets when they need more challenging benchmarks. A similar situation happens with computer vision datasets such as MNIST and ImageNet, which have been widely explored. The state of the art models essentially solves these problems, motivating the search for more challenging tasks. Therefore, this paper introduces a new computer vision toy dataset featuring iris flowers. Users of a nature photography application took the pictures, thus they include noisy background information. Additionally, certain desirable features are not guaranteed, such as single, similarly-sized objects at the center of each picture, which makes the task more challenging. Our benchmark results show that the dataset can be challenging for traditional machine learning algorithms without any pre-processing steps, while the state of the art deep learning architectures achieve around 82% accuracy, which means some effort will be necessary to drive this accuracy closer to what has been accomplished for MNIST and ImageNet.

BRACIS 2021

Iris-CV: Classifying Iris Flowers Is Not as Easy as You Thought

machine learning

computer vision

dataset

**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

Several coarsening algorithms have been developed as a powerful strategy to deal with difficult machine learning problems represented by large-scale networks, including, network visualization, trajectory mining, community detection and dimension reduction. It iteratively reduces the original network into a hierarchy of gradually smaller informative representations. However, few of these algorithms have been specifically designed to deal with bipartite networks and they still face theoretical limitations that need to be explored. Specifically, a recently introduced algorithm, called MLPb, is based on a synchronous label propagation strategy. In spite of an interesting approach, it presents the following two problems: 1) A high-cost search strategy in dense networks and 2) the cyclic oscillation problem yielded by the synchronous propagation scheme. In this paper, we address these issues and propose a novel fast coarsening algorithm more suitable for large-scale bipartite networks. Our proposal introduces a semi-synchronous strategy via cross-propagation, which allows a time-effective implementation and deeply reduces the oscillation phenomenon. The empirical analysis in both synthetic networks and real-world networks shows that our coarsening strategy outperforms previous approaches regarding accuracy and runtime.

Coarsening Algorithm via Semi-Synchronous Label Propagation for Bipartite Networks

As changes in external environments are inevitable, a lifelong mapping system is desirable for autonomous robots that aim at long-term operation. Capturing external environment changes into internal representations (for example, maps) is crucial for proper behavior and safety, especially in the case of autonomous vehicles. In this work, we propose a new large-scale mapping system for our autonomous vehicle or any other. The new mapping system is based on the Graph SLAM algorithm, with extensions to deal with the calibration of odometry directly in the optimization of the graph and to address map merging for long-term map maintenance. The mapping system can use sensor data from one or more robots to build and merge different types of occupancy grid maps. The system's performance is evaluated in a series of experiments carried out with data captured in complex real-world scenarios. The experimental results indicate that the new large-scale mapping system can provide high-quality occupancy grid maps for later navigation and localization of autonomous vehicles.

Long-term Map Maintenance in Complex Environments

In this paper, we present a new corpus for Rhetorical Role Identification in Portuguese legal documents. The corpus comprises petitions from 70 civil lawsuits filed in TJMS court and was manually labeled with rhetorical roles specifically tailored for petitions. Since petition documents are created without a standard structure, we had to deal with several issues to clean the extracted textual content. We assessed classic and deep learning machine learning methods on the proposed corpus.The best performing method obtained a F-score of 80.50. At the best of our knowledge, this is the first work to deal with rhetorical role identification for petitions, given that previous works focused only on judicial decisions. Additionally, it is also the first work to tackle this task for the Portuguese language. The proposed corpus, as well the proposed rhetorical roles, can foster new research in the judicial area and also lead to new solutions to improve the flow of Brazilian court houses.

Rhetorical Role Identification for Portuguese Legal Documents

Improving the interaction between consumers and marketplaces,
focusing on reaching higher conversion rates is one of the main
goals of e-commerce companies. Offering better results for user queries is
mandatory to improve user experience and convert it into purchases. This
paper investigates how named entity recognition can extract relevant attributes
from product titles to derive better lters for user queries. We
conducted several experiments based on MITIE and BERT applied to
smartphones/cellphones product titles from the largest Brazilian retail ecommerce.
Both of our strategies achieve outstanding results with a general
F1 score of around 95%.We concluded that using a classical machine
learning pipeline is still more useful than relying on large pre-trained language
models, considering the model's throughput and efficiency. Future
work may focus on evaluating the scalability and reusability capacity of
both approaches.

Named Entity Recognition for Brazilian Portuguese Product Titles

In the different areas of knowledge, textual data are important sources of information. This way, Information Extraction methods have been developed to identify and structure information present in textual documents. In particular there is the Named Entity Recognition (NER) task, which consists of using methods to identify Named Entities, such as Person, Place, among others, in texts, using techniques from Natural Language Processing and Machine Learning. Recent works explored the use of external sources of knowledge to boost the Machine Learning models with sets of domain specific relevant information for the NER task. This work aims to evaluate the aggregation of external knowl edge, in the form of Gazetter and Knowledge Graphs, for NER task. Our approach is composed of two steps: i) generation of embeddings, ii) defi nition and training of the Machine Learning methods. The experiments were conducted on four English datasets, and their results show that the applied strategies for external knowledge integration did not bring great gains to the models, as expressed by F1-Score metric. In the performed experiments, there was an F1-score increase in 17 of the 32 cases where external knowledge was used, but in most cases the gains were lesser than 0.5% in F1-score. In some scenarios the aggregated external knowledge does not capture relevant content, thus not being necessarily beneficial to the methodology.

When external knowledge does not aggregate in Named Entity Recognition

Competitive intelligence (CI) is a relevant area of a corporation and can support the strategic business area by showing those responsible, helping decision-making on how to position an organization in the market. This work uses the Bidirectional Transformer Encoding Representations (BERT) to process a sentence and its named entities and extract the parts of the sentences that represent or describe the semantic relationship between these named entities. The approach was developed for the Portuguese language, considering the financial domain and exploring deep linguistic representations without using other lexical-semantic resources. The results of the experiments show a precision of 73.5\% using the Jaccard metric that measures the similarity between sentences. The second contribution of this work is the manually constructed dataset with more than 4.500 tuples (phrase, entity, entity) annotated.

Entity Relation Extraction from News Articles in Portuguese for Competitive Intelligence based on BERT

Deep learning models have been the state-of-the-art for a variety of challenging tasks in natural language processing, but to achieve good results they often require big labeled datasets. Deep active learning algorithms were designed to reduce the annotation cost for training such models. Current deep active learning algorithms, however, aim at training a good deep learning model with as little labeled data as possible, and as such are not useful in scenarios where the full dataset must be labeled. As a solution to this problem, this work investigates deep active-self learning algorithms that employ self-labeling using the trained model to help alleviate the cost of annotating full datasets for named entity recognition tasks. The experiments performed indicate that the proposed deep active-self learning algorithm is capable of reducing manual annotation costs for labeling the complete dataset for named entity recognition with less than 2\% of the self labeled tokens being mislabeled. We also investigate an early stopping technique that doesn't rely on a validation set, which effectively reduces even further the annotation costs of the proposed active-self learning algorithm in real world scenarios.

Deep Active-Self Learning Applied to Named Entity Recognition

The increasing influence from users in social media has made that aggressive content disseminates over the internet. To tackle this problem, recent advances in Aggressive Language Detection have demonstrated a good performance of Deep Learning techniques. Recently Transformer based architectures such as Bidirectional Encoder Representations from Transformer (BERT) outperformed previous aggressive text detection baselines. However, most of the Transformers-based approaches are unable to properly capture global information such as language vocabulary. Thus, in this work, we focus on aggressive content detection using the combination of Vocabulary Graph Convolutional Network (VGCN) to capture global information and BERT to model local information. This combined approach called VGCN-BERT allows us to improve the feature level representation in Spanish aggressive language detection. Our experiments were performed on a benchmark called MEX-A3T aggressiveness dataset which is composed of aggressive and non-aggressive Tweets written in the Mexican Spanish variant. We report 86.46\% in terms of F1-score using this VGCN-BERT approach which allows us to obtain comparable results with the current state-of-the-art, ensemble BERT, so as to detect aggressive content regarding the track MEX-A3T 2020.

Aggressive Language Detection using VGCN-BERT for Spanish Texts

Studies in sentiment analysis have examined how different features are effective in identifying subjective sentences. Several studies on sentiment analysis exist in the literature that have already performed some evaluation of NLP classifications. However, the vast majority of them did not handle texts in the Brazilian Portuguese language, and there is no one to consider the combination of sets of text features of NLP tasks with classifiers. Therefore, in our investigation, we combined empirical features to identify subjective sentences in Portuguese and provide a comprehensive analysis of each set of features’ relative importance using a representative set of user reviews.

An Empirical Study of Text Features for Identifying Subjective Sentences in Portuguese

Indexing artwork is not only a tedious job; it is an impossible task to complete manually given the amount of online art. In any case, the automatic classification of art styles is also a challenge due to the relative lack of labeled data and the complexity of the subject matter. This complexity means that common data augmentation techniques may not generate useful data; in fact, they may degrade performance in practice. In this paper, we use Generative Adversarial Networks for data augmentation so as to improve the accuracy of an art style classifier, showing that we can improve performance of EfficientNet B0, a state of art classifier. To achieve this result, we introduce Class-by-Class Performance Analysis; we also present a modified version of the SAGAN training configuration that allows better control against mode collapse and vanishing gradient in the context of artwork.

Premium content

Iris-CV: Classifying Iris Flowers Is Not as Easy as You Thought

Downloads

Next from BRACIS 2021

Coarsening Algorithm via Semi-Synchronous Label Propagation for Bipartite Networks

Similar lecture

Classification of chest X-ray images using Machine Learning and Histogram of Oriented Gradients