http://hdl.handle.net/10174/1135 2026-04-04T00:12:43Z 2026-04-04T00:12:43Z Blom, Berry Pereira, L. M. Pereira http://hdl.handle.net/10174/40225 2026-01-07T22:11:41Z 2023-11-15T00:00:00Z

Title: Domain Adaptation in Transformer Models: Question Answering of Dutch Government Policies Authors: Blom, Berry; Pereira, L. M. Pereira Editors: Quaresma, Paulo; Camacho, David; Yin, Hujun; Gonçalves, Teresa; Julian, Vicente; Tallón-Ballesteros, Antonio J. Abstract: Automatic answering questions helps users in finding information efficiently, in contrast with web search engines that require keywords to be provided and large texts to be processed. The first Dutch Question Answering (QA) system uses basic natural language processing techniques based on text similarity between the question and the answer. After the introduction of pre-trained transformer-based models like BERT, higher scores were achieved with over 7.7% improvement for the General Language Understanding Evaluation (GLUE) score. Pre-trained transformer-based models tend to over-generalize when applied to a specific domain, leading to less precise context-specific outputs. There is a marked research gap in experiment strategies to adapt these models effectively for domain-specific applications. Additionally, there is a lack of Dutch resources for automatic question answering, as the only existing dataset, Dutch SQuAD, is a translation of the SQuAD dataset in English. We propose a new dataset, PolicyQA, containing questions and answers about Dutch government policies and use domain adaptation techniques to address the generalizability problem of transformer-based models. The experimental setup includes the Long Short-Term memory (LSTM), a baseline neural network, and three BERT-based models, mBert, RobBERT, and BERTje, with domain adaptation. The datasets used for testing are the proposed PolicyQA dataset and the existing Dutch SQuAD. From the results, we found that the multilanguage BERT-model, mBert, outperforms the Dutch BERT-based models (RobBERT and BERTje) on the both datasets. By introducing fine-tuning, a domain adaptation technique, the mBert model improved to 94.10% of F1-score, a gain of 226% compared to its performance without fine-tuning.

2023-11-15T00:00:00Z Lamar-Leon, Javier Alonso Baryolo, Raul Salgueiro, Pedro Garcia Reyes, Edel Gonzalez Diaz, Rocio http://hdl.handle.net/10174/36939 2024-06-01T15:44:10Z 2023-11-27T00:00:00Z

Title: Abandoned Object Detection Using Persistent Homology Authors: Lamar-Leon, Javier; Alonso Baryolo, Raul; Salgueiro, Pedro; Garcia Reyes, Edel; Gonzalez Diaz, Rocio Editors: Vasconcelos, Verónica; Domingues, Inês; Paredes, Simão Abstract: The automatic detection of suspicious abandoned objects has become a priority in video surveillance in the last years. Terror- ist attacks, improperly parked vehicles, abandoned drug packages and many other events, endorse the interest in automating this task. It is challenge to detect such objects due to many issues present in public spaces for video-sequence process, like occlusions, illumination changes, crowded environments, etc. On the other hand, using deep learning can be difficult due to the fact that it is more successful in perceptual tasks and generally what are called system 1 tasks. In this work we propose to use topological features to describe the scenery objects. These features have been used in objects with dynamic shape and maintain the stability under perturbations. The objects (foreground) are the result of to apply a background subtraction algorithm. We propose the concept the surveil- lance points: set of points uniformly distributed on scene. Then we keep track of the changes in a cubic region centered at each surveillance points. For that, we construct a simplicial complex (topological space) from the k foreground frames. We obtain the topological features (using per- sistent homology) in the sub-complexes for each cubical-regions, which represents the activity around the surveillance points. Finally for each surveillance points we keep track of the changes of its associated topo- logical signature in time, in order to detect the abandoned objects. The accuracy of our method is tested on PETS2006 database with promising results

2023-11-27T00:00:00Z Medeiros, Eduardo Corado, Leonel Rato, Luis Quaresma, Paulo Salgueiro, Pedro http://hdl.handle.net/10174/34466 2023-02-15T15:56:42Z 2022-09-30T23:00:00Z

Title: Optimized European Portuguese Speech-To-Text using Deep Learning Authors: Medeiros, Eduardo; Corado, Leonel; Rato, Luis; Quaresma, Paulo; Salgueiro, Pedro Abstract: We have developed an ASR system for European Portuguese implement ing the QuartzNet [3] architecture with the NeMo [4] framework. Two approaches were used in this work: from scratch and using transfer learning. The experiments were data-driven focused instead of algorithm finetuning. Experiments confirm that models developed using transfer learning have shown better results (WER=0.0513) than developing models from scratch (WER=0.1945).

2022-09-30T23:00:00Z Tereso, Marco Rato, Luis Gonçalves, Teresa http://hdl.handle.net/10174/34094 2023-02-10T11:24:25Z 2020-05-31T23:00:00Z

Title: Automatic classification of ornamental stones using Machine Learning techniques - a study applied to limestone. Authors: Tereso, Marco; Rato, Luis; Gonçalves, Teresa Abstract: The industry of extraction and transformation of rock minerals has an enormous importance in the Portuguese trade balance. The export volume increases every year, and to maintain these results it is necessary to invest in the modernization and optimization of production processes, as well as, in the classification of raw materials. This study aims to implement a classification model of ornamental rocks through the analysis and classification of images, using machine learning algorithms. The recognition of the type of stone, through the capture of images and subsequent algorithmic analysis, will allow to define quality control scales in future processes, taking into account the different types of stone. In addition, it will also allow to develop models capable of helping in reducing the amount of raw material wasted. This work presents the steps taken to create a classification model, using a dataset of 2260 images distributed over four classes, three of which are very similar to color level and one with a different tone. In this study, the results of the application of three automatic classification algorithms are analyzed. In addition, a discussion of how types of images can improve results and the execution times of algorithms are presented.

2020-05-31T23:00:00Z