João Graça


Co-founder and CTO of Unbabel
A Scalable Consistent High Quality Translation Pipeline
Abstract

Despite recent developments on machine translation its actual quality is still very low and only usable for gisting purposes. On the other hand human translation can provide the desired quality levels but tends to be too slow and very process driven to achieve the required scale. In this talk I will present a pipeline that combines the best of both worlds using different AI modules to achieve high throughput translation with consistently high quality and its use cases in industry.

Fernando Pereira

VP and Engineering Fellow at Google
Representation learning, inference, and reasoning
Abstract
Advances in deep learning have led to a golden age of increasingly rich models of language with large experimental gains in practical language understanding tasks. However, these gains came at the expense of structured inference for global constraint satisfaction and multistep reasoning.  I will illustrate this with examples that are obvious for people but pose unsolved inference and reasoning challenges to current ML methods. I will conclude with questions and proposed tasks that may sharpen our understanding of these issues.

Cícero dos Santos

Research Scientist at at Amazon
Natural Language Processing in the Age of Deep Representation Learning
Abstract
The ability to develop effective Natural Language Processing (NLP) systems free of handcrafted features has made of Deep Learning the most successful story in recent NLP research. At the core of this success is the use of deep distributed representations, which can be automatically learned in both supervised and self-supervised ways. In this talk, I will present some of my recent contributions on deep representation learning for NLP. These contributions are focused on novel neural network architectures, loss functions and training approaches, which have pushed the state-of-the-art in different tasks such as named entity recognition, sentiment analysis, question answering and text summarization.

Isabel Trancoso

Senior Researcher
Spoken Language Systems Lab
INESC-ID / IST, Univ. Lisbon, Portugal
Speech – a health biomarker
Abstract
The potential of speech as a biomarker for health has been realized for diseases affecting respiratory organs, such as the common Cold, or Obstructive Sleep Apnea, for mood disorders such as Depression, and Bipolar Disease, and neurodegenerative diseases such as Parkinson’s, Alzheimer’s, and Huntington’s disease. The recent progress achieved with ML methods trained with lab data recorded in very controlled conditions has been impressive, but it also raises questions of robustness and privacy: how do these methods scale to in-the-wild data? What is their dependence on language and demographic factors? How can diagnosis be performed preserving the privacy of the speaker? I will conclude with a vision of speech analysis being as common as a blood test for diagnosis and monitoring of speech affecting diseases.