Tackling Climate Change with Machine Learning NeurIPS 2020 Virtual - - PowerPoint PPT Presentation

Tackling Climate Change with Machine Learning NeurIPS 2020 Virtual Workshop

December 11-12, 2020

Why topic detection for climate change?

2

Documents Sentiment analysis Question-answering Fact-checking Output

Tackling Climate Change with Machine Learning

• Some examples:

Not as easy as it seems…

Tackling Climate Change with Machine Learning

CLIMATEXT

4

Tackling Climate Change with Machine Learning

Wikipedia data (Wiki-Doc-Train/Dev/Test)

5

Tackling Climate Change with Machine Learning

Positives Wikipedia Links Graph

1) Land surfaces are heating faster than the

• cean surface, leading to

heatwaves, wildfires, and the expansion of deserts. 2) As the temperature difference between the Arctic and the equator decreases, ocean currents that are driven by that temperature difference, like the Gulf Stream. 3) …. Documents .csv

Weakly labeling

Negatives

AL-Wiki and AL-10Ks train data

6

Tackling Climate Change with Machine Learning

10-Ks Wikipedia

10-Ks (unlabeled)

AL - WIKI AL - 10-Ks

NB model AL-WIKI NB model AL-10-Ks

Wikipedia, 10K, and Claims evaluation data

7

Tackling Climate Change with Machine Learning

Document label sampling scheme (labeled by raters) BERT-predictions sampling scheme (labeled by raters) Websites of claim collections and other (labeled by raters)

Main Result

8

Our contributions:

1.

We introduce CLIMATEXT, a dataset for sentence-based climate change topic detection, which we make publicly available.

2.

We analyze keyword-based, naïve-Bayesian, and a BERT-based approach to explore their performance in identifying climate-change relevant text in Wikipedia, 10-K filings, and in climate-related claims database. Going forward:

• Make the annotated data public.
• Improve algorithms to detect climate-change topic in a wide range of text sources.

Tackling Climate Change with Machine Learning