RESUMO
Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder in the world, and it is characterized by the production of different motor and non-motor symptoms which negatively affect speech and language production. For decades, the research community has been working on methodologies to automatically model these biomarkers to detect and monitor the disease; however, although speech impairments have been widely explored, language remains underexplored despite being a valuable source of information, especially to assess cognitive impairments associated with non-motor symptoms. This study proposes the automatic assessment of PD patients using different methodologies to model speech and language biomarkers. One-dimensional and two-dimensional convolutional neural networks (CNNs), along with pre-trained models such as Wav2Vec 2.0, BERT, and BETO, were considered to classify PD patients vs. Healthy Control (HC) subjects. The first approach consisted of modeling speech and language independently. Then, the best representations from each modality were combined following early, joint, and late fusion strategies. The results show that the speech modality yielded an accuracy of up to 88%, thus outperforming all language representations, including the multi-modal approach. These results suggest that speech representations better discriminate PD patients and HC subjects than language representations. When analyzing the fusion strategies, we observed that changes in the time span of the multi-modal representation could produce a significant loss of information in the speech modality, which was likely linked to a decrease in accuracy in the multi-modal experiments. Further experiments are necessary to validate this claim with other fusion methods using different time spans.
RESUMO
Resources for Natural Language Processing (NLP) are less numerous for languages different from English. In the clinical domain, where these resources are vital for obtaining new knowledge about human health and diseases, creating new resources for the Spanish language is imperative. One of the most common approaches in NLP is word embeddings, which are dense vector representations of a word, considering the word's context. This vector representation is usually the first step in various NLP tasks, such as text classification or information extraction. Therefore, in order to enrich Spanish language NLP tools, we built a Spanish clinical corpus from waiting list diagnostic suspicions, a biomedical corpus from medical journals, and term sequences sampled from the Unified Medical Language System (UMLS). These three corpora can be used to compute word embeddings models from scratch using Word2vec and fastText algorithms. Furthermore, to validate the quality of the calculated embeddings, we adapted several evaluation datasets in English, including some tests that have not been used in Spanish to the best of our knowledge. These translations were validated by two bilingual clinicians following an ad hoc validation standard for the translation. Even though contextualized word embeddings nowadays receive enormous attention, their calculation and deployment require specialized hardware and giant training corpora. Our static embeddings can be used in clinical applications with limited computational resources. The validation of the intrinsic test we present here can help groups working on static and contextualized word embeddings. We are releasing the training corpus and the embeddings within this publication.
RESUMO
The Internet's popularization has increased the amount of content produced and consumed on the web. To take advantage of this new market, major content producers such as Netflix and Amazon Prime have emerged, focusing on video streaming services. However, despite the large number and diversity of videos made available by these content providers, few of them attract the attention of most users. For example, in the data explored in this article, only 6% of the most popular videos account for 85% of total views. Finding out in advance which videos will be popular is not trivial, especially given many influencing variables. Nevertheless, a tool with this ability would be of great value to help dimension network infrastructure and properly recommend new content to users. In this way, this manuscript examines the machine learning-based approaches that have been proposed to solve the prediction of web content popularity. To this end, we first survey the literature and elaborate a taxonomy that classifies models according to predictive features and describes state-of-the-art features and techniques used to solve this task. While analyzing previous works, we saw an opportunity to use textual features for video prediction. Thus, additionally, we propose a case study that combines features acquired through attribute engineering and word embedding to predict the popularity of a video. The first approach is based on predictive attributes defined by resource engineering. The second takes advantage of word embeddings from video descriptions and titles. We experimented with the proposed techniques in a set of videos from GloboPlay, the largest provider of video streaming services in Latin America. A combination of engineering features and embeddings using the Random Forest algorithm achieved the best result, with an accuracy of 87%.