ABSTRACT. Event extraction is an Information Retrieval task that com- monly consists of identifying the central word for the event (trigger) and the event’s arguments. This task has been extensively studied for English but lags behind for Portuguese, partly due to the lack of task- specific annotated corpora. This paper proposes a framework in which two separated BERT-based models were fine-tuned to identify and clas- sify events in Portuguese documents. We decompose this task into two sub-tasks. Firstly, we use a token classification model to detect event triggers. To extract event arguments, we train a Question Answering model that queries the triggers about their corresponding event argu- ment roles. Given the lack of event annotated corpora in Portuguese, we translated the original version of the ACE-2005 dataset (a reference in the field) into Portuguese, producing a new corpus for Portuguese event extraction. To accomplish this, we developed an automatic translation pipeline. Our framework obtains F1 marks of 64.4 for trigger classifica- tion and 46.7 for argument classification setting, thus a new state of the art reference for these tasks in Portuguese.

ABSTRACT. The recent success of Large Language Models (LLM) has sparked concerns about their potential to spread misinformation. As a result, there is a pressing need for tools to identify "fake arguments" generated by such models. To create these tools, examples of texts generated by LLMs are needed. This paper introduces a methodology to obtain good, bad and ugly arguments from argumentative essays produced by ChatGPT, OpenAI's LLM. We then describe a novel dataset containing a set of diverse arguments, ArGPT. We assess the effectiveness of our dataset and establish baselines for several argumentation-related tasks. Finally, we show that the artificially generated data relates well to human argumentation and thus is useful as a tool to train and test systems for the defined tasks.

ABSTRACT. Argumentation Mining (AM) is a growing sub-field of Natural Language Processing (NLP) which aims at extracting argumentative structures from text. In this work, neural learning and symbolic reasoning are combined in a system named N-SAUR that extracts the argumentative structures present in a collection of texts and then assesses each argument's strength. The extraction is based on the Toulmin's model and the result quality surpasses previous approaches over an existing benchmark. Complementary scores are also extracted and combined with a set of rules that produce the final calculation of argument strength. The performance of the system was evaluated through human assessments. Users can also interact with the system in various ways, allowing for the strength calculation to change through user-cooperative reasoning.

ABSTRACT. To advance the neural encoding of Portuguese (PT), and a fortiori the technological preparation of this language for the digital age, we developed a Transformer-based foundation model that sets a new state of the art in this respect for two of its variants, namely European Portuguese from Portugal (PT-PT) and American Portuguese from Brazil (PT-BR).

To develop this encoder, which we named Abertina PT-*, a strong model was used as a starting point, DeBERTa, and its pre-training was done over data sets of Portuguese, namely over a data set we gathered for PT-PT and over the brWaC corpus for PT-BR. The performance of Abertina and competing models was assessed by evaluating them on prominent downstream language processing tasks adapted for Portuguese.

Both Abertina PT-PT and PT-BR versions are distributed free of charge and under the most permissive license possible and can be run on consumer-grade hardware, thus seeking to contribute to the advancement of research and innovation in language technology for Portuguese.

ABSTRACT. We describe OSPT, a new linguistic resource for European Portuguese that comprises more than 1.5 million Portuguese-Portuguese sentential paraphrase pairs. We generated the pairs automatically by using neural machine translation to translate the non-Portuguese side of a large parallel corpus. We hope this new corpus can be a valuable resource for paraphrase generation and provide a rich semantic knowledge source to improve downstream natural language understanding tasks. To show the quality and utility of such a dataset, we use it to train paraphrastic sentence embeddings and evaluate them in the ASSIN2 semantic textual similarity (STS) competition. We found that semantic embeddings trained on a small subset of OSPT can produce better semantic embeddings than the ones trained in the finely curated ASSIN2's training data. Additionally, we show OSPT can be used for paraphrase generation with the potential to produce good data augmentation systems that pseudo-translate from Brazilian Portuguese to European Portuguese.

ABSTRACT. E-commerce has become an essential aspect of modern life, providing consumers worldwide with convenience and accessibility. However, the high volume of short and noisy product descriptions in text streams of massive e-commerce platforms translates into an increased number of clusters, presenting challenges for standard model-based stream clustering algorithms. This is the case of a dataset extracted from the Brazilian NF-e Project containing electronic invoice product descriptions, including many product clusters. While LDA-based clustering methods have shown to be crucial, they have been mainly evaluated on datasets with few clusters. We propose the Topic Model with Contextual Outlier Handling (TMCOH) method to overcome this limitation. This method combines the Dirichlet Process, specific word representation, and contextual outlier detection techniques to recycle identified outliers aiming to integrate them into appropriate clusters later on. The experimental results for our case study demonstrate the effectiveness of TMCOH when compared to state-of-the-art methods and its potential for application to text clustering in large datasets.

ABSTRACT. Identifying small dolphin species based on their vocalizations remains a challenging task due to their similar vocal signatures and frequency modulation patterns, particularly when the available data sets are relatively limited. To address this issue, a new feature set has been introduced that focuses on capturing both the predominant frequency range of the vocalizations and other higher level details in the spectral contour, which are valuable for distinguishing between small dolphin species. These features are computed from two distinct representations of the vocalizations: the short time Fourier transform and Mel frequency cepstral coefficients. By utilizing these features with two popular classifiers (K-Nearest Neighbors and Support Vector Machines), a model accuracy of 95.47% has been achieved, representing an improvement over previous studies.

ABSTRACT. Probabilistic supervised learning within the Bayesian paradigm typically use Gaussian Processes (GPs) to model the sought function, and provide a means for securing reliable uncertainties in said functional learning, while offering interpretability. Prediction of the output of such a learnt function is closed-form in this approach. In this work, we present GP-based learning of the functional relation between two variables, using various kinds of kernels that are called in to parametrise the covariance function of the invoked GP. However, such covariance kernels are typically parametric in the literature, with hyperparameters that are learnt from the data. Here, we discuss a new nonparametric covariance kernel, and compare its performance against existing non-stationary and stationary kernels, as well as against Deep Neural Networks. We present results on both univariate and multivariate data, to demonstrate the range of applicability of the presented learning scheme.

ABSTRACT. This work explores pre-training as a strategy to allow reinforcement learning (RL) algorithms to efficiently adapt to new ---albeit similar--- tasks. We argue for introducing variability during the pre-training phase, in the form of augmentations to the observations of the agent, to improve the sample-efficiency of the fine-tuning stage. We propose to categorize such variability in the form of perceptual, dynamic and semantic augmentations, which can be easily employed in standard pre-training methods. We perform extensive evaluations of our proposed augmentation scheme in model-based algorithms, across multiple scenarios of increasing complexity. The results consistently show that our augmentation scheme significantly improves the efficiency of the fine-tuning to novel tasks, outperforming other state-of-the-art pre-training approaches.

ABSTRACT. Learning to forecast spatiotemporal environmental processes is relatively understudied within the context of robotics. While there is prior work in deep learning for spatiotemporal process learning in domains such as high-frequency trading and video surveillance, these works often investigate learning from a set of timesteps where each timestep contains the entire spatial context. Such approaches may not be useful when we can only gather sparse amounts of data at each timestep. An example of this is robotic sampling for information gathering, such as using UAVs/UGVs for weather monitoring or animal tracking. In this work, we propose a methodology that leverages a neural methodology called Recurrent Neural Processes to learn spatiotemporal environmental dynamics for forecasting from selective samples gathered by a team of robots using a mixture of Gaussian Processes model in an online learning fashion. Thus, we combine two learning paradigms in that we use an active learning approach to adaptively gather informative samples and a supervised learning approach to capture and predict complex spatiotemporal environmental phenomena.

ABSTRACT. In recent years, the increasing availability of distributed data has led to a growing interest in transfer learning across multiple nodes. However, local data may not be adequate to learn sufficiently accurate models, and the problem of learning from multiple distributed sources remains a challenge. To address this issue, Machine Learning Data Markets (MLDM) have been proposed as a potential solution. In MLDM, autonomous agents exchange relevant data in a cooperative relationship to improve their models. Previous research has shown that data exchange can lead to better models, but this has only been demonstrated with only two agents. In this paper, we present an extended evaluation of a simple version of the MLDM framework in a collaborative scenario. Our experiments show that data exchange has the potential to improve learning performance, even in a simple version of MLDM. The findings conclude that there exists a direct correlation between the number of agents and the gained performance, while an inverse correlation was observed between the performance and the data batch sizes. The results of this study provide important insights into the effectiveness of MLDM and how it can be used to improve learning performance in distributed systems. By increasing the number of agents, a more efficient system can be achieved, while larger data batch sizes can decrease the global performance of the system. These observations highlight the importance of considering both the number of agents and the data batch sizes when designing distributed learning systems using the MLDM framework.

ABSTRACT. Quadruped robots are gaining attention in the research community because of their superior mobility and versatility in a wide range of applications. However, they are restricted to procedures that do not need precise object interaction. With the addition of a robotic arm, they can overcome these drawbacks and be used in a new set of tasks. Combining a legged robot's dextrous movement with a robotic arm's maneuverability allows the emergence of a highly flexible system, with the disadvantage of higher complexity of motion planning and control methods. This paper gives an overview of the existing quadruped systems capable of manipulation, with a particular interest in systems with high movement flexibility. The main topics discussed are the motion planning approaches and the selected kinematic configuration. This review concludes that the most followed research path is to add a robotic arm on the quadrupedal base and that the motion planning approach used depends on the desired application. For simple tasks, the arm can be seen as an independent system, which is simpler to implement. For more complex jobs the coupling effects between the arm and quadruped robot must be considered.

ABSTRACT. The progress of Natural Language Processing (NLP), although fast in recent years, is not at the same pace for all languages. African languages in particular are still behind and lack automatic processing tools. Some of these tools are very important for the development of these languages but also have an important role in many NLP applications. This is particularly the case for automatic spell checkers. Several approaches have been studied to address this task and the one modeling spelling correction as a translation task from misspelled (noisy) text to well-spelled (correct) text shows promising results. However, this approach requires a parallel corpus of noisy data on the one hand and correct data on the other hand, whereas Wolof is a low-resource language and does not have such a corpus. In this paper, we present a way to address the constraint related to the lack of data by generating synthetic data and we present sequence-to-sequence models using Deep Learning for spelling correction in Wolof. We evaluated these models in three different scenarios depending on the subwording method applied to the data and showed that the latter had a significant impact on the performance of the models, which opens the way for future research in Wolof spelling correction.

ABSTRACT. Deep learning approaches have become popular in many different areas including sentiment analysis (SA), because of their competitive performance. The downside of this approach is that they do not provide understandable explanations on how the sentiment values are calculated. In contrast, previous approaches that used sentiment lexicons for sentiment analysis tasks can do that, but their performance is normally not high. To leverage the strengths of both approaches, we present a neuro-symbolic approach that combines symbolic and deep learning (DL) methods for SA tasks. The first one exploits sentiment lexicon and shifter patterns. The DL approach uses a pretrained language model (PLM) to construct sentiment lexicon. Our experimental results show that the proposed approach leads to promising results. Although the results did not reach the level of the DL approach, a great advantage is that sentiment prediction can be accompanied by understandable explanations

ABSTRACT. Topics discussed on social media platforms contain a dis- parate amount of information written in colloquial language, making it difficult to understand the narrative of the topic. In this paper, we take a step forward, towards the resolution of this problem by proposing a framework that performs the automatic extraction of narratives from a document, such as tweet posts. To this regard, we propose a method- ology that extracts information from the texts through a pipeline of tasks, such as co-reference resolution and the extraction of entity rela- tions. The result of this process is embedded into an annotation file to be used by subsequent operations, such as visualization schemas. We named this framework Tweet2Story and measured its effectiveness under an evaluation schema that involved three different aspects: (i) as an Open Information extraction (OpenIE) task, (ii) by comparing the narratives of manually annotated news articles linked to tweets about the same topic and (iii) by comparing their knowledge graphs, produced by the narratives, in a qualitative way. The results obtained show a high pre- cision and a moderate recall, on par with other OpenIE state-of-the-art frameworks and confirm that the narratives can be extracted from small texts. Furthermore, we show that the narrative can be visualized in an easily understandable way.

ABSTRACT. Dialogue systems need to deal with the unpredictability of user intents to track dialogue state and the heterogeneity of slots to understand user preferences. In this paper we investigate the hypothesis that solving these challenges as one unified model will allow the transfer of parameter support data across the different tasks. The proposed principled model is based on a Transformer encoder, trained on multiple tasks, and leveraged by a rich input that conditions the model on the target inferences. Conditioning the Transformer encoder on multiple target inferences over the same corpus, i.e., intent and multiple slot types, allows learning richer language interactions than a single-task model would be able to. In fact, experimental results demonstrate that conditioning the model on an increasing number of dialogue inference tasks leads to improved results: on the MultiWOZ dataset, the joint intent and slot detection can by improved by 3.2\% by conditioning on intent, 10.8\% by conditioning on slot and 14.4\% by conditioning on both intent and slots. Moreover, on real conversations with Farfetch users, we observed an improvement of 10\% when used a conditioned BERT.

ABSTRACT. This study examines how perturbations in image, audio, and text inputs affect the performance of different classification models. Various perturbators were applied to three seed datasets at different intensities to produce noisy test data. Then, the models' performance was evaluated on the generated test data. The findings indicate that there is a consistent relationship between larger perturbations and lower model performance across perturbators, models, and domains. However, this relationship varies depending on the characteristics of the specific model, dataset, and perturbator.

ABSTRACT. Attention-based agents have had much success in many areas of Arti-ficial Intelligence, such as Deep Reinforcement Learning. This work revisits two such architectures, namely, Deep Attention Recurrent Q-Networks (DARQNs) and Soft Top-Down Spatial Attention (STDA) and explores the similarities between them. More specifically, this work tries to improve the per-formance of the DARQN architecture by leveraging elements proposed by the STDA architecture, such as the formulation of its attention function which also includes the incorporation of a spatial basis into its computation. The imple-mentation tested, denoted Deep Attention Recurrent Actor-Critic (DARAC), uses the A2C learning algorithm. The results obtained seem to suggest that the incorporation of the spatial basis can improve performance in some cases. Somewhat surprisingly, the revised formulations of the attention function tested provided little to no improvement. Overall, DARAC showed competitive results when compared to STDA and slightly better in some of the experiments. All the experiments were validated using the Atari 2600 videogame benchmark.

ABSTRACT. Neural networks have achieved remarkable success in various applications such as image classification, speech recognition, and natural language processing. However, the growing size of neural networks poses significant challenges in terms of memory usage, computational cost, and deployment on resource-constrained devices. Pruning is a popular technique to reduce the complexity of neural networks by removing unnecessary connections, neurons, or filters. In this paper, we present novel pruning algorithms that can reduce the number of parameters in neural networks by up to 98\% without sacrificing accuracy. This is done by scaling the pruning rate of the models to the size of the model and scheduling the pruning to execute throughout the training of the model.

ABSTRACT. The widespread of distributed renewable energy is leading to an increased need for advanced energy management solutions in buildings. The variability of generation needs to be balanced by consumer flexibility, which needs to be accomplished by keeping the consumption cost as low as possible, while guaranteeing consumer comfort. This paper proposes a rule-based system with the aim of generating recommendations for actions regarding the energy management of different energy consumption devices, namely lights and air conditioning. The proposed set of rules considers the forecasted values of building generation, consumption, user presence in different rooms and energy prices. In this way, building energy management systems are endowed with increased adaptability and reliability considering the lowering of energy costs and maintenance of user comfort. Results, using real data from an office building, demonstrate the appropriateness of the proposed model in generating recommendations that are in line with current context.

ABSTRACT. One big challenge of evolutionary computation techniques, and especially genetic programming, is the computational power that is needed to use them. One solution to address this issue is using distributed computation techniques to distribute the workload to speed up the learning process. The combination of genetic programming with federated learning could solve the computational distribution while promoting a collaborative learning environment. The use of federated learning allows several clients to create a federation where they can create and train a collaborative model, with the benefit of not sharing personal data. This paper proposes a federated learning configuration that enables the use of genetic programming for its global model. In addition, this paper also proposes a new aggregation algorithm that enables the collaborative evolution of genetic programming individuals in federated learning. The case study uses flexible genetic programming, an existing and successful algorithm for image classification, integrated into a federated learning framework. The results show the benefits of combining genetic programming with federated learning.

ABSTRACT. Energy load disaggregation can contribute to balancing power grids by enhancing the effectiveness of demand-side management and promoting electricity saving behavior through increased consumer awareness. However, the field currently lacks a comprehensive overview. To address this gap, this paper conducts a scoping review of load disaggregation domains, data types, and methods, by assessing 72 full-text journal articles. The findings reveal that domestic electricity consumption is the most researched area, while others, such as industrial load disaggregation, are rarely discussed. The majority of research uses relatively low-frequency data, sampled between 1 and 60 seconds. A wide variety of methods are used, and artificial neural networks are the most common, followed by optimization strategies, Hidden Markov Models, and Graph Signal Processing approaches.

ABSTRACT. With the remnants of a COVID-19 pandemic still crippling the European economy, and the Russo-Ukrainian war propagating this crisis even further, it has become more than crucial to invest in renewable energy resources to mitigate energy dependencies. As a result, these crises have lowered the competitiveness of European manufacturers when compared to the rest of the world. Nevertheless, ma-chine longevity is also essential to consider in manufacturing environments, since maintenance costs due to poor load management can lead to considerable additional monetary costs in the long term. The premise of the present paper is to propose a production scheduling algorithm that focuses on optimizing the total energy costs and machine longevity in a flexible job shop manufacturing layout. To achieve this, a Genetic Algorithm is employed to shift tasks in order to reduce load during peak demand times, utilize locally generated energy to its potential, minimize single-machine task overload, and consider imposed constraints in the production schedule. To validate the proposed methodology, a case study from the literature that uses real-production data is explored and compared to the present paper’s solution. Results show that the proposed methodology was capable of reducing single-machine task overload, that is, improving machine longevity, by 87.8%, while only increasing the energy costs, as a consequence, by 12.8%.

ABSTRACT. This paper tackles a single-machine scheduling problem where each job is characterized by weight, duration, due date, and deadline, while the objective is to minimize the weighted number of tardy jobs. The problem is known to be strongly NP-hard and has practical applications in various domains, such as customer service and production planning. The best known exact approach uses a branch-and-bound structure, but its efficiency varies depending on the distribution of job parameters. To address this, we propose a new data-driven heuristic algorithm that considers the parameter distribution and uses machine learning and integer linear programming to improve the optimality gap. The algorithm also guarantees to obtain a feasible solution if it exists. Experimental results show that the proposed approach outperforms the current state-of-the-art heuristic.

ABSTRACT. Learning a well-informed heuristic function for hard planning domains is an elusive problem. Although there are known neural network architectures to represent such heuristic knowledge, it is not obvious what concrete information is learned and whether techniques aimed at understanding the structure help in improving the quality of the heuristics.

This paper presents a network model that learns a heuristic function capable of relating distant parts of the state space via optimal plan imitation using the attention mechanism which drastically improves the learning of a good heuristic function. To counter the limitation of this method in the creation of problems of increasing difficulty, we demonstrate the use of curriculum learning, where newly solved problem instances are added to the training set, which, in turn, helps to solve problems of higher complexities and far exceeds the performances of all existing baselines including classical planning heuristics. We demonstrate its effectiveness on grid-type PDDL domains.

ABSTRACT. The aim of this paper is to summarize agent-based evacuation models in Python by conducting a systematic literature search using the PRISMA methodology. The principles of evacuation models are briefly described. Python packages and libraries for agent-based modelling frameworks are explained. Two research questions are defined. The first question aims to find out what a typical agent-based evacuation model looks like, the second question focuses on the details of the use of the Python programming language. The results of the review process are presented. Overall, Python is a suitable language for the development of agent-based evacuation models, as evidenced by the number of programming li-braries and tools, as well as the growing number of scientific publications. How-ever, most of the currently published models suffer from many shortcomings. The most surprising is the lack of an ODD protocol and source codes.