NLP for Energy Transition - Demo Day

October 2020.

Transcript

1. Using NLP and sentiment analysis to better interpret and implement energy transition 1 Source:- https://www.vox.com/2019/4/22/18510953/climate-change-2020-bernie-sanders-donald-trump-harvard-poll
2. What to do with all this data? How do we analyze all this information? Challenges • Multiple data sources. • Incomplete information. • How do we account for biases? • What role do citizens play? • Languages and finances. The Problem – Lack of Social Interpretation • https://www.worldenergy.org/news- views/entry/global-challenge-to-define-a-role- of-a-citizen-in-energy-transition 2 Source:- https://www.vox.com/2019/4/22/18510953/climate-change-2020-bernie- sanders-donald-trump-harvard-poll
3. Purpose & Motivation, Data-The New Oil • A shift in support for alternatives requires social upheaval and greater understanding among populations. • Consumers and individuals vote with their wallets and must be educated on the merits and demerits of each strategy. • It is thus imperative to interpret social sentiments regarding existing and emerging sources as well as technologies globally. • Data is the fuel that can facilitate a better decision making system for environmentalists and policy makers. 3
4. Those being studied: • Reddit, Facebook, Youtube, Twitter. • Weibo. • News Websites like Huffington Post, Reuters, Bloomberg etc. • Developing and developed nations, including those poised to be global superpowers. • U.S., China, India, Nigeria, Europe etc. Problem Scope 4 Those being benefitted: • Policy makers and economists. • Government stakeholders and environmentalists. • Consumers, tax payers and society.
5. • List data sources from social media sources, forums, news articles and surveys related to energy and energy transition. • Analyze and breakdown the data to remove noise. • Create visualizations to help understand social sentiments across nations and societies for various issues including energy prices and comfortability to switch sources. The Goal 5 OBJECTIVES WHAT WE ACHIEVED • Cleaned datasets with proper labels from various media sources. • Interpretable sentiment analysis visualizations for specific keywords and phrases. • Highly accurate AI models to predict and classify texts. • Automation code to collect, scrape and analyze text from social media.
6. 6 PIPILINE STEPS Find relevant links and discussion forums Connect website data and scrape sources Automate data cleaning and filtering relevant content Perform sentiment analysis and compare best results Build code to predict and classify texts
7. 7 BREAKING DOWN DATA- SENTIMENT ANALYSIS • Data from online sources is usually unclean, filled with irrelevant text, numbers, articles, stop words and whitespace. • Data has to be segmented, tokenized, vectorized and categorized by comparing it with existing dictionaries with surplus information about words and their context. • Data cleanup for NLP is also essential to computing speeds since less processing space is wasted.
8. BREAKING DOWN THE TASKS 8 Collect Data Clean Data Generate Features Convert data to usable forms Load Dictionaries Generate Sentiments Create bag of words Generate tabular features Train and test models for prediction
9. 9 TOPIC MODELLING AND CLUSTERING
12. Sentiment Analysis • Text analysis method • Can detect polarity (positive or negative opinion) • Can be very precise (fine-grained) • Can be aspect-based • Can detect emotion 12
13. 13 Source: Devopedia SENTIMENT ANALYSIS STRATEGIES
14. 14 BIGRAMS AND TRIGRAMS
15. 15 SAMPLE FREQUENCY PLOTS
16. 16 DICTIONARY BASED ANALYSIS
17. 17 SAMPLE RESULTS-REDDIT
18. EDA of Tweets Collocated wordclouds Visualizing co-occuring words for different countries 18
19. India Nigeria 19 https://omdena.com/projects/ai-energy/
20. UK USA 20
21. EDA of Tweets Co-occurrence Networks Visualizing top 25 bigrams as networks 21
22. India USA 22
23. UK Nigeria 23
24. Transfer Learning • Use a pre-trained model →« Portability » of features → When small amount of data • Permitted breakthrough of NLP 24
25. In pratice How we applied it for • Text classification • Graph Analytics 25
26. Transfer Learning on Tweets Text Classification 26
27. • General • Focus on costs Two Fine-Tuned Models 27
28. • Pre-trained model (BERT base) • BookCorpus • Wikipedia English dataset • Fine-tuned on scraped tweets • Positive • Negative 98% accuracy General 28
29. On their opinions about energy transition Comparing Countries 29
30. • Same pre-trained model • Fine-tuned on Climate-44 dataset • With sentiment index (between 0 and 1) • Threshold 0.5 94% accuracy Focus On Costs 30
31. + Clustering on claims Classify Unseen Tweets Talking About Costs 31
32. Community detection • Method for finding groups within complex systems • Represented on a graph 32
33. Transfer Learning on Tweets Graph Analytics (aspect based) 33
34. How renewable energy and its related topics traverse around USA and the UK 34
35. Steps (1) • Map tweets with Open Street Map data (location) • Identiy topics by applying topic model • Filter tweets on renewable energies 35
36. Steps (2) • Use fine-tuned BERT sentiment classifier • Label remaining using Label Propagation • Perform community detection → show how an aspect-based sentiment traverses around in nearby locations 36
37. Climate Change Topic 37
40. Renewable energy topic 40
43. Topic traversal in Florida Sentiment traversal in Florida 43
45. Key Findings From EDA to models outputs 45
46. • Topic variation by geographical area and country • Ex : electricty availability isn’t discussed everywhere • Discussions forums and articles can talk about same concerns • Sentiments flowing on nearby cities • Some countries with an overall feeling more positive than others 46
47. Improvements 47
48. • Apply graph approach to other areas • Scalable code base • See how fine-tuned models performs on any commentts (not only Twitter) 48
49. 49 METHOD COMPARISONS
50. 50 THE PRESENT AND THE ROAD AFAR • We have a repository of information about social sentiments( different countries, websites, languages). • Sentiment analysis results for sources other than English Spanish and Chinese(Mandarin, Cantonese). • Current generated models for topic synthesis and prediction are well adjusted to dissect information obtained from new social media sources. • Ensemble methods like LinearSVC are better at connecting sentiments to context. • Social media is a rich resource to tap into and understand what people are talking about. • Social perceptions about energy bills, energy use, green energy and more topics vary depending on online anonymity and platform culture. • Reddit and Facebook are preferred choices for social sentiment analysis. • Sentiment analysis needs to expands beyond text(pictures, emojis etc).