Sentence simplification (SS) focuses on adapting sentences to enhance their readability and accessibility. While large language models (LLMs) match task-specific baselines in English SS, their performance in Portuguese remains underexplored. This paper presents a comprehensive performance comparison of 26 state-of-the-art LLMs in Portuguese SS, alongside two simplification models trained explicitly for this task and language. They are evaluated under a one-shot setting across scientific, news, and government datasets. We benchmark the models with our newly introduced Gov-Lang-BR corpus (1,703 complex-simple sentence pairs from Brazilian government agencies) and two established datasets: PorSimplesSent and Museum-PT. Our investigation takes advantage of both automatic metrics and large-scale linguistic analysis to examine the transformations achieved by the LLMs. Furthermore, a qualitative assessment of selected generated outputs provides deeper insights into simplification quality. Our findings reveal that while open-source LLMs have achieved impressive results, closed-source LLMs continue to outperform them in Portuguese SS.

As the number of fine-tuned language models for specialized domains and tasks continues to grow, managing a diverse set of solutions presents increasing challenges regarding scalability and adaptability. In this context, model-fusion techniques offer a promising approach to enhancing generalization by leveraging knowledge from multiple trained models. This paper investigates the fusion of pre-trained BERT-based models using ten different fusion strategies: (i) Simple Merging, (ii) Select Simple Merging, (iii) Fisher Merging, (iv) Select Fisher Merging, (v) RegMean, (vi) Task Arithmetic, (vii) DARE-simple Merging, (viii) TIES-MERGING, (ix) Robust Fine-tuning, and (x) Select Epoch Merging. To assess the effectiveness of these techniques, we conducted experiments in the challenging task of detecting hate speech in Portuguese. Such a task benefits from combining knowledge with model fusion, given that individual models might overlook the context-dependent nature of offensive language and nuanced forms of hate speech. The results using six datasets indicate that TIES-MERGING, in particular, can outperform individual models by successfully integrating specialized knowledge into a single solution, a more efficient and robust model.

Traditional neural networks (NNs) learn primarily from data, which limits their capacity to represent relational knowledge or handle symbolic relational data effectively. Although graph neural networks (GNNs) address this limitation at the level of relational data, they continue to struggle at learning relational knowledge. Neural-symbolic learning offers a solution by combining machine learning with knowledge representation, enabling the development of interpretable logic-based models learned from neural networks. Bottom clause propositionalization (BCP) is a prominent approach that transforms relational knowledge into attribute-value examples. A bottom clause is a logical representation created from each example as a starting point for the search process. BCP can be used with symbolic learners or neural networks to tackle relational domains. However, BCP often faces significant memory storage problems when handling larger datasets due to the volume of logical literals that it generates. Semi-propositionalization can alleviate these storage problems by grouping logical literals. However, it does not eliminate the substantial time requirements to create a bottom clause for each example. This paper investigates the application of sampling to the examples used for bottom clause generation. The hypothesis is that the number of examples needed to generate bottom clauses can be reduced significantly. Finding representative bottom clauses from data should enable relational learning to take place at an adequate level of abstract relational knowledge rather than simply at the level of the relations between any two data points. We evaluate this hypothesis by training a classifier with different sampling from five relational datasets. We experimentally validate the size of each sampling for each dataset. Experimental results show that training classifiers with fewer relational examples produces competitive results compared to using the entire dataset. The best results are obtained with up to 50% reduction in the set of examples.

Recommender systems play a crucial role in helping users decide what to watch or purchase by suggesting relevant items.These systems can enhance the media experience by considering user preferences, inferring behavior, and personalized recommendations.However, users often do not understand why a particular item was recommended to them.Explainable recommender systems aim to clarify the reasoning behind recommendations, increasing user trust and confidence.Despite advancements, gaps remain in the literature, particularly in evaluating these systems.This thesis will explore and propose new metrics to better assess explanation methods, investigating why and how current explanations fall short in evaluations.Additionally, we aim to examine whether explanations can reveal if recommender systems create filter bubbles and explore ways to mitigate this issue based on user preferences.

This paper proposes a multimodal retrieval system for Chest CT (called ChestFinder) that combines image and report embeddings’ into a filter-and-query strategy. ChestFinder is composed of three modules, namely (i) text transformation, (ii) feature extraction, and (iii) ranking aggregation. Text transformation is conducted by a fine-tuned Generative Pre-training Transformer model (GPT), and the image embeddings are extracted after (i) training a Residual Neural Network (ResNet-50), and (ii) reducing and scaling the encoded vectors. ChestFinder produces one list of similar images and another of related reports for each query input composed of a Chest CT with a radiology report. Then, the ChestFinder ranking aggregation module fuses those two lists to produce the final ordered set of retrieved objects, as in a top-k query. The aggregation is performed by a fine-tuned Threshold Algorithm (TA) whose weights are calculated by a Multi-Layer Perceptron (MLP) trained to label the reports. To examine the quality enhancement brought by this multimodal search, we constructed a dataset of Chest CTs from our University Hospital PACS/RIS systems by filtering distinct cases diagnosed with emphysema (one finding per case and with at least two radiologists agreeing on the diagnosis). A holdout experimental evaluation showed the ChestFinder search achieved higher Accuracy and Sensitivity than content-only top-k searches. Results also indicated quality gains drawn from the adjustments of ChestFinder modular components: (i) fine-tuned GPT achieved up to 0.89 F1-Score in data testing with a stable train/validation ratio for radiology reports, (ii) fine-tuned GPT significantly outperformed the zero-shot approach as well as a fine-tuned BERT, (iii) non-weighted ranking aggregation increased the search accuracy in up to 10%, and (iv) fine-tuned TA outperformed the baseline and non-weighted ranking aggregation in up to 52%.

With the popularization of video streaming services, it has become increasingly important to discover which videos will be popular to prepare the network infrastructure. Popularity prediction has been studied with several Machine Learning models and with different features captured from videos. This article presents a multimodal model that uses different classifiers fed by different features, building a robust and flexible model that surpasses models used in practice. To build this model, we used data from Globoplay, the largest streaming service in Latin America. On the other hand, predicting content popularity from a catalog of available media can identify videos that demand more resources from the network infrastructure, allowing service providers to adopt preventive measures to maintain transmission quality. Notably, we analyze if visual features extracted from thumbnails add value to this task. We experiment with the proposed approaches on a set of videos from GloboPlay. Our model gives the best accuracy found, in addition to the advantages of robustness and flexibility, adapting better to practical cases of popularity prediction.

Abstract Text style transfer (TST) aims at automatically changing a text’s stylistic features, such as formality, sentiment, authorial style, humor, and complexity, while still trying to preserve its content. Although the scientific community has investigated TST since the 1980s, it has recently regained attention by adopting deep unsupervised strategies to address the challenge of training without parallel data. In this manuscript, we investigate how relying on sequence-to-sequence pretraining models affects the performance of TST when the pretraining step leverages pairs of paraphrase data. Furthermore, we propose a new technique to enhance the sequence-to-sequence model by distilling knowledge from masked language models. We evaluate our proposals on three unsupervised style transfer tasks with widely used benchmarks: author imitation, formality transfer, and polarity swap. The evaluation relies on quantitative and qualitative analyses and comparisons with the results of state-of-the-art models. For the author imitation and the formality transfer task, we show that using the proposed techniques improves all measured metrics and leads to state-of-the-art (SOTA) results in content preservation and an overall score in the author imitation domain. In the formality transfer domain, we paired with the SOTA method in the style control metric. Regarding the polarity swap domain, we show that the knowledge distillation component improves all measured metrics. The paraphrase pretraining increases content preservation at the expense of harming style control. Based on the results reached in these domains, we also discuss in the manuscript if the tasks we address have the same nature and should be equally treated as TST tasks.

Major Depression Disorder (MDD) is a mental condition that causes severe impairments in a person's life. Early detection of such a condition is imperative, especially to detect cases where the person does not even know that they are affected by it. Social media provide a ubiquitous platform with self-generated data that Machine Learning models enhanced with Natural Language Processing mechanisms can extensively explore. Nonetheless, we advocate that previous Machine Learning approaches to early detection of MDD do not adequately model the problem at hand. Most assume that the isolated publications on social media are enough for early detection. This paper proposes a new problem formulation for this task using the multiple-instance learning paradigm. Furthermore, we include a theoretical and experimental analysis of the method on a dataset of Brazilian university students using the transformer and LSTM architectures. The results indicate that the proposed approach yields better accuracy and explanation capabilities than previous studies that do not rely on multiple-instance learning.

Depression has increased at alarming rates in the worldwide population. One alternative to finding depressed individuals is using social media data to train machine learning (ML) models to identify depressed cases automatically. Previous works have already relied on ML to solve this task with reasonably good F-measure scores. Still, several limitations prevent the full potential of these models. In this work, we show that the depression identification task through social media is better modeled as a Multiple Instance Learning (MIL) problem that can exploit the temporal dependencies between posts.

Summary Most scientific experiments can be modeled as workflows. These workflows are usually computing‐ and data‐intensive, demanding the use of high‐performance computing environments such as clusters, grids, and clouds. This latter offers the advantage of the elasticity, which allows for changing the number of virtual machines (VMs) on demand. Workflows are typically managed using scientific workflow management systems (SWfMS). Many existing SWfMSs offer support for cloud‐based execution. Each SWfMS has its scheduler that follows a well‐defined cost function. However, such cost functions should consider the characteristics of a dynamic environment, such as live migrations or performance fluctuations, which are far from trivial to model. This article proposes a novel scheduling strategy, named ReASSIgN, based on reinforcement learning (RL). By relying on an RL technique, one may assume that there is an optimal (or suboptimal) solution for the scheduling problem, and aims at learning the best scheduling based on previous executions in the absence of a mathematical model of the environment. For this, an extension of a well‐known workflow simulator WorkflowSim is proposed to implement an RL strategy for scheduling workflows. Once the scheduling plan is generated via simulation, the workflow is executed in the cloud using SciCumulus SWfMS. We conducted a throughout evaluation of the proposed scheduling strategy using a real astronomy workflow named Montage.

Arguably, player behavior profiling is one of the most relevant tasks of Game Analytics. However, to fulfill the needs of this task, gameplay data should be handled so that the player behavior can be profiled and even understood. Usually, gameplay data is stored as raw log-like files, from which gameplay metrics are computed. However, gameplay metrics have been commonly used as input to classify player behavior with two drawbacks: (1) gameplay metrics are mostly handcrafted and (2) they might not be adequate for fine-grain analysis as they are just computed after key events, such as stage or game completion. In this paper, we present a novel approach for player profiling based on provenance graphs, an alternative to log-like files that model causal relationships between entities in game. Our approach leverages recent advances in deep learning over graph representation of player states and its neighboring contexts, requiring no handcrafted features. We perform clustering on learned nodes representations to profile at a fine-grain the player behavior in provenance data collected from a multiplayer battle game and assess the obtained profiles through statistical analysis and data visualization.

Disorder of the thyroid glands is a widespread health problem. Early detection of thyroid cancer increases the chances of effective treatment. Dynamic infrared thermal imaging (DITI) is an examination technique that has been recently studied and applied for investigation and diagnosis of different diseases. Patterns allowing differentiate regions of malignant from benignant nodules is a task of great importance in DITI. In this work, two techniques of analysis of thyroid nodules with infrared thermography are investigated: the use of a single thermogram and the use of temperature series. The images used are available for public use by other researchers, and the used techniques are completely described as well as the achieved results. No other works using infrared images until now have considered DITI examination for thyroid nodules investigation. Moreover, it is the first work to release the used images for public use and possible future comparison.

Smart cities initiatives have the potential to improve the life of citizens in a huge number of dimensions. One of them is the development of techniques and services capable of contributing to the enhancement of security public policies. Finding criminal patterns from historical data would arguably help in predicting and even preventing thefts and burglaries that continuously increase in urban centers worldwide. However, accessing such history and finding patterns across the interrelated crime occurrences data are challenging tasks, particularly to underdevelopment countries. In this paper, we address these problems by combining three techniques: we collect crime data from existing crowd-sourcing systems, we automatically induce patterns with relational machine learning, and we manage the entire process using scientific workflows. The framework developed under these lines is named CRiMINaL (Crime patteRn MachINe Learning). Experimental results conducted from a popular Brazilian source of data and a traditional relational learning system shows that CRiMINaL is a promising tool to induce interpretable models that can assist police departments on crime prevention.

Automated story generation is a desired feature in games and interactive media because it can control how a virtual world evolves so that it can be adapted to the players' choices. In order to have variety and quality in the generated stories, previous works have relied on simulation‐based storytelling, in which a story is generated as their characters, represented as agents, try to achieve their goals. One challenge of this approach is to make the agents act more like human characters and less like omniscient intelligent beings. In this article, we present a perception model for simulation‐based story generation that introduces errors into characters' knowledge, (mis)leading them to non‐optimal, but still coherent, believable actions. The perception is executed using a description of the virtual world's elements using physical characteristics, and a pattern matching process that associates combinations of physical characteristics with predefined combinations of attributes, which are allowed to be wrong, and consequently may result in non‐perfect interpretations of the world. We developed a story generation system from the proposed model and tested it with a version of the Little Red Riding Hood story, famous for its perception failure. Our results show interesting variations for the traditional known ending.

FORTE_MBC is a First-Order Logic theory revision system, built upon the FORTE system but with an important supplement: it makes use of (i) a Bottom Clause to define the search space of literals and (ii) a set of mode declarations to validate the yielded clauses. Introducing the Bottom Clause was essential to reduce the runtime of the revision process (experimental results showed an average speed-up of 55|$\times$|⁠), since one of the key steps of the revision process is to create literals to add to clauses. However, the effectiveness and efficiency of learning and revising as a refinement process heavily rely on other factors, such as the generality relation that induces a generalization model for the search space, and, as a consequence, the refinement operators. These components have never been formally analysed in FORTE(_MBC). In this work, we contribute with (i) an adaptation of the existing theoretical frameworks that characterize refinement operators to define the search space and revision operators of systems like FORTE_MBC, which use a set of Bottom Clauses and mode declarations to constrain the search; (ii) an improvement of the theory refinement operators of FORTE_MBC to make them ideal for the defined space. We present the feasibility of these modified operators by implementing them in the FORTE_MBC system. Experimental results show that we are indeed able to obtain a more efficient revision process by using the proposed ideal operators.

Deep learning models have recently achieved the state-of-the-art results on a well-known pedestrian detection dataset. However, such images were obtained from open scenarios with fixed imaging geometry parameters, which may produce a network not suitable for detecting a person in more general settings, such as the ones found in surveillance systems. As gathering and annotating data is a highly expensive manual task, we propose a methodology for artificially augmenting the positive training set with automatically generated local image affine and perspective transforms. Furthermore, to enrich the variability of background images, we include to the negative training set images that resemble human figures automatically obtained by the proposed methodology over images from commonly found surveillance scenarios. Extensive results show that by providing the enriched data as the input to a Convolutional Neural Network it is possible to precisely detect pedestrians in a number of public datasets. The data enrichment proposed here may also be used in other detectors based on supervised learning architectures, as the process is independent from the learning algorithm employed.