Learning to Predict the Global Risks Interconnections from the Web [ - PowerPoint PPT Presentation

Learning to Predict the Global Risks Interconnections from the Web [ Minerva: AI/ML for News ] Dr. Ernesto Diaz-Aviles Co-Founder, CEO and Chief Scientist at Libre AI and Adjunct Assistant Professor at UCD <ernesto@libreai.com> https://libreai.com 2018-08-27 Supported by Google through the Digital News Initiative

Dr. Ernesto Diaz-Aviles Co-Founder and CEO Scientist and Engineer with 15 years of experience deploying AI, ML, and Data-driven solutions at scale. Former Chief Data Scientist and VP at Citi’s Innovation Lab. Research Scientist at IBM Research. Research Fellow at the Web Science Lab, L3S Research Center, Germany. Claudia Orellana-Rodriguez, M.Sc. Co-Founder and Chief Scientific Officer Claudia is a scientist and engineer whose work leverages the power of machine learning, natural language processing, social network analysis, and opinion mining to unveil patterns of engagement, attention and influence on the digital era. Claudia is also a researcher at the Insight Centre for Data Analytics in UCD and collaborator with the MIT Center for Civic Media. 2

Our mission is to widely disseminate the benefits of Artificial Intelligence and Machine Learning and make them accessible to the world 3

AI and News We envision a future where journalists will no longer be limited to report past or current affairs , but they will be empowered by Artificial Intelligence to write about future events with a fair degree of certainty 4

AI and News ● Everything is connected and there are clear historical signs and cycles that produce very similar consequences. The understanding of such interconnections and causality is fundamental for a comprehensive news coverage ● However, connecting the dots and discovering the multiple relationships among events, entities, and global risks are not trivial tasks for journalists 5

World Economic Forum: The Global Risks Interconnections A " global risk " is defined as an uncertain event or condition that, if it occurs, can cause significant negative impact for several countries or industries within the next 10 years. The Global risks 5 broader classes: (1) Economic Risks (2) Environmental Risks (3) Geopolitical Risks (4) Societal Risks (5) Technological Risks 6 WEF: https://www.weforum.org/reports/the-global-risks-report-2018

Minerva: automatically generate a Global Risks Interconnections Map from large news datasets and web sources 7

Project: Minerva Learning to Predict the Global Risks Interconnections from the Web Prototype based on Artificial Intelligence and Machine Learning that mines the Web and predicts the (non-obvious) interconnections of global risks that will be at the core of tomorrow's news 8

Minerva: Learn to Predict the Global Risks Interconnections from Data <<discover>> Libre AI for News - Classification of news articles into Global Risks minerva - Detection of key entities : persons, organizations, locations - Unveil existing relationships : graph of <<enhance>> interconnections - Predict: infer future connections 9

Minerva: AI/ML Pipeline Common Crawl News Daily: ~ 4GB - 5GB In 2018 ~ 1.5 T http://commoncrawl.org/2016/ 10/news-dataset-available/ Global Risk Classification Entity Extraction Relation Extraction Nowcasting Global Risk Graph Creation Visualization 10

Minerva: AI/ML Pipeline no Stream of Risk / No-Risk Documents Global Risk Classification Entity Extraction Document Global Risk Prediction Embedding Risk Classifier Relation Extraction Nowcasting yes Global Risk Graph Creation Visualization 11

Minerva: AI/ML Pipeline no Stream of Risk / No-Risk Documents Global Risk Classification Entity Extraction Document Global Risk Prediction Embedding Risk Classifier Relation Extraction Article Extraction - News Please: Nowcasting yes https://github.com/fhamborg/news-please - Unicode, Dammit Global Risk https://www.crummy.com/software/Beauti Graph Creation fulSoup/bs4/doc/ Visualization Embedding GloVe 12 https://nlp.stanford.edu/projects/glove/

Minerva: AI/ML Pipeline Stream of Documents Global Risk (entity, global risk) Classification Entity Extractor Entity Extraction Relation Extraction Nowcasting Global Risk Graph Creation Visualization 13

Minerva: AI/ML Pipeline Stream of Documents Global Risk (entity, global risk) Classification Entity Extractor Entity Extraction Relation Extraction Nowcasting Global Risk NLP – NER Graph Creation SpaCy: Embed, encode, attend, predict. CNN + GloVe Visualization https://spacy.io/ 14

Minerva: AI/ML Pipeline Stream of Documents (entity, global risk) Global Risk Relation Extraction Global Risks Links Classification Entity Extraction Relation Extraction Nowcasting Global Risk Graph Creation Visualization 15

Minerva: AI/ML Pipeline Relation Extraction Stream of (entity, global risk) Global Risks Links Documents Global Risk risk_i risk_j Classification w_ij Entity Extraction Relation Strategy 1: Jaccard Similarity (faster) Extraction |entities_i ∩ entities_j| Nowcasting w_ij = ------------------------- |entities_i ∪ entities_j| Global Risk Graph Creation Strategy 2: Semantic Similarity Visualization w_ij = sim ( embedding (entities_i), embedding (entities_j)) 16

Minerva: AI/ML Pipeline Stream of Documents t_1 t_2 t_n t_n+1 Global Risk Classification ... Entity Extraction Relation Extraction extract predict Nowcasting Global Risk Graph Creation Visualization 17

Minerva: AI/ML Pipeline t_1 t_2 t_n t_n+1 Stream of ... Documents Global Risk Classification extract predict Entity Extraction Relation risk_i risk_j Extraction w_ij Nowcasting Global Risk Graph Creation w_ij_t1, w_ij_t2, … w_ij_tn -> w_ij_t_n+1 Visualization 18 extract predict

Minerva: AI/ML Pipeline t_1 t_2 t_n t_n+1 ... Stream of Documents Global Risk Classification extract predict Entity Extraction Time Series Forecasting with CNN - Conv1D with dilation and "causal" padding Relation risk_i risk_j - ~ WaveNet Extraction w_ij - All link weights predicted simultaneously Nowcasting - Keras / TensorFlow https://keras.io/layers/convolutional/ Global Risk Graph Creation w_ij_t1, w_ij_t2, … w_ij_tn -> w_ij_t_n+1 Visualization 19 extract predict

Minerva: Interactive Graph Stream of Documents Global Risk Classification Entity Extraction Relation Extraction Nowcasting Global Risk Graph Creation Interactive Visualization Visualization D3.js https://d3js.org/ 20

War without Rules Offensive cyber capabilities are developing more rapidly than our ability to deal with hostile incidents. This creates a fog of uncertainty in which potential miscalculations could trigger a spiral of retaliatory responses. Imagine that a country’s critical infrastructure systems are compromised by a cyberattack, leading to disruption of essential services and loss of life—the pressure to retaliate would build rapidly, potentially setting off an escalatory chain reaction. [...] 21 WEF: https://www.weforum.org/reports/the-global-risks-report-2018

Experimental Evaluation ● Are the main connections between Global Risks predicted? ● Ground truth: Web Economic Forum reports ● Metric (averaged over all risks): |{relevant links} ∩ {top-n predicted links}| Precision@n = ------------------------------------------- n ● Dataset Commons Crawl News 2018 sample (all articles are in English), Articles from Irish News media outlets, Major press agencies, and Major newspapers around the world 22

Experimental Evaluation 2018 23

Conclusion ● “ It is difficult to make predictions, especially about the future ” – Danish Proverb ● Predicted Global Risk Interconnection more accurate for short horizons (Nowcasting). E.g., months or quarters rather than 10 years as WEF definition ● Entity based relations are promising proxy for risk interconnections ● Computationally cheaper Jaccard similarity leads to better precision than embedding-based strategy ● Next: continue evaluation and User study ● Initial rollout with partners. If interested, let me know 24

AI and News We envision a future where journalists will no longer be limited to report past or current affairs , but they will be empowered by Artificial Intelligence to write about future events with a fair degree of certainty Thank you! Dr. Ernesto Diaz-Aviles Co-Founder, CEO and Chief Scientist at Libre AI and Adjunct Assistant Professor at UCD <ernesto@libreai.com> 25 https://libreai.com

LibreAI.com 26