3 November 2021 in Computational Linguistics
Word embeddings are vectorial semantic representations built with either counting or predicting techniques aimed at capturing shades of meaning from word co-occurrences. Since their introduction, these representations have been criticized for lacking interpretable dimensions. This property of word embeddings limits our understanding of the semantic features they actually encode. Moreover, it contributes to the “black box” nature of the tasks in which they are used, since the reasons for word embedding performance often remain opaque to humans. In this contribution, we explore the semantic properties encoded in word embeddings by mapping them onto interpretable vectors, consisting of explicit and neurobiologically motivated semantic features (Binder et al. 2016). Our exploration takes into account different types of embeddings, including factorized count vectors and predict models (Skip-Gram, GloVe, etc.), as well as the most recent contextualized representations (i.e., ELMo and BERT).In our analysis, we first evaluate the quality of the mapping in a retrieval task, then we shed light on the semantic features that are better encoded in each embedding type. A large number of probing tasks is finally set to assess how the original and the mapped embeddings perform in discriminating semantic categories. For each probing task, we identify the most relevant semantic features and we show that there is a correlation between the embedding performance and how they encode those features. This study sets itself as a step forward in understanding which aspects of meaning are captured by vector spaces, by proposing a new and simple method to carve human-interpretable semantic representations from distributional vectors.
1 November 2020 in Journal of Pain and Symptom Management
12 March 2021 in Journal of Medical Internet Research
Artificial intelligence (AI) technologies can play a key role in preventing, detecting, and monitoring epidemics. In this paper, we provide an overview of the recently published literature on the COVID-19 pandemic in four strategic areas: (1) triage, diagnosis, and risk prediction; (2) drug repurposing and development; (3) pharmacogenomics and vaccines; and (4) mining of the medical literature. We highlight how AI-powered health care can enable public health systems to efficiently handle future outbreaks and improve patient outcomes.
1 June 2023 in Knowledge-Based Systems
14 January 2021 in npj Digital Medicine
Targeted contact-tracing through mobile phone apps has been proposed as an instrument to help contain the spread of COVID-19 and manage the lifting of nation-wide lock-downs currently in place in USA and Europe. However, there is an ongoing debate on its potential efficacy, especially in light of region-specific demographics. We built an expanded SIR model of COVID-19 epidemics that accounts for region-specific population densities, and we used it to test the impact of a contact-tracing app in a number of scenarios. Using demographic and mobility data from Italy and Spain, we used the model to simulate scenarios that vary in baseline contact rates, population densities, and fraction of app users in the population. Our results show that, in support of efficient isolation of symptomatic cases, app-mediated contact-tracing can successfully mitigate the epidemic even with a relatively small fraction of users, and even suppress altogether with a larger fraction of users. However, when regional differences in population density are taken into consideration, the epidemic can be significantly harder to contain in higher density areas, highlighting potential limitations of this intervention in specific contexts. This work corroborates previous results in favor of app-mediated contact-tracing as mitigation measure for COVID-19, and draws attention on the importance of region-specific demographic and mobility factors to achieve maximum efficacy in containment policies.