OPINION MINING CLIMATE CHANGE TWEETS

By: Lily Lou, Lizzie Tong, Michael Yang, Haocheng Zhao

Introduction

• 1.5 degrees Celsius between 2030 and 2052 if greenhouse gas emissions (GHG) continue at current rate
• More intense weather, ecosystems destroyed, crises over scarcity over water and natural resources, migration due to climate change
• Predict viewpoints of climate change Tweets and through exploratory analysis, find common topics when climate change is mentioned

Related Work

Pak, Alexander and Paroubek, Mucha. “Twitter as a Corpus for Sentiment Analysis and Opinion Mining,” 2010. https://lexitron.nectec.or.th/public/LREC-2010_Malta/pdf/385_Paper.pdf.

Neethu, M S, and R Rajasree. “Sentiment Analysis in Twitter Using Machine Learning Techniques.” In 2013 Fourth International Conference on Computing, Communications and Networking Technologies (ICCCNT), 1–5, 2013. https://doi.org/10.1109/ICCCNT.2013.6726818.

​

​

​

​

Related Work

Mucha, Nithisha. “Sentiment Analysis of Global Warming Using Twitter Data,” 2018. https://library.ndsu.edu/ir/handle/10365/28166.

Reyes-Menendez, Ana, José Ramón Saura, and Cesar Alvarez-Alonso. “Understanding #WorldEnvironmentDay User Opinions in Twitter: A Topic-Based Sentiment Analysis Approach.” International Journal of Environmental Research and Public Health 15, no. 11 (November 2018): 2537. https://doi.org/10.3390/ijerph15112537.

​

​

​

​

Exploratory Topic Analysis

• Nvivo Pro 12 - Word Frequency

​

• Greater than 50 instances

​

• Including stem-words/ synonyms to proxy topics/arguments across tweets

CLIMATE CHANGE BELIEVERS

​

​

CLIMATE CHANGE DENIERS

Predicting Sentiment Analysis

• What patterns exist within features that indicate whether or not the author believes in climate change?
• What features are particularly accurate?

​

• Looked at different combinations of features:
• Unigram, bigram, trigram, count occurences, stemming, line length
• Compare Results!

​

Sentiment Analysis: Approaches

Approaches: Model

• Built Naive Bayes model
• Previous literature indicated that Naive Bayes was superior to other methods such as Support Vector Machine (SVM), Conditional Random Fields, Logistic Regression, etc.
• Given smaller number of instances, evaluated using Cross Validation

​

Methodology: Dataset & Cleansing

Tweets with keywords “climate change” scraped from Twitter (Credit to IBM employees)

6090 in total. Columns:Tweet, existence, existence.confidence

Watson labeled each tweet with negative or positive attitude towards climate change

What a Tweet in our dataset looks like?

“RT @mmfa: Brain Freeze: Conservative media still using winter weather to attack global warming: http://bit.ly/9nKEcc #p2 #noisemachine #climate”

“"D.C. Snowstorm: How Global Warming Makes Blizzards Worse - TIME" ( http://bit.ly/9sMGXj ) It's snowing because it's so warm? Cahmaaahn.(via @TonyMackGD) [link]”

A lot of info we don’t need in a tweet.

“RT @mmfa: Brain Freeze: Conservative media still using winter weather to attack global warming: http://bit.ly/9nKEcc #p2 #noisemachine #climate”

“"D.C. Snowstorm: How Global Warming Makes Blizzards Worse - TIME" ( http://bit.ly/9sMGXj ) It's snowing because it's so warm? Cahmaaahn. (via @TonyMackGD) [link]”

​

There are something we are not certain...

“RT @mmfa: Brain Freeze: Conservative media still using winter weather to attack global warming: http://bit.ly/9nKEcc #p2 #noisemachine #climate”

“"D.C. Snowstorm: How Global Warming Makes Blizzards Worse - TIME" ( http://bit.ly/9sMGXj ) It's snowing because it's so warm? Cahmaaahn. (via @TonyMackGD) [link]”

​

AND…. ALL the punctuations are unnecessary

“RT @mmfa: Brain Freeze: Conservative media still using winter weather to attack global warming: http://bit.ly/9nKEcc #p2 #noisemachine #climate”

“"D.C. Snowstorm: How Global Warming Makes Blizzards Worse - TIME" ( http://bit.ly/9sMGXj ) It's snowing because it's so warm? Cahmaaahn. (via @TonyMackGD) [link]”

​

After cleaning process

“well the truth is brian we can t solve global warming because i fucking changed light bulbs in my house barack obama”

“un process in danger unless world agrees on climate change telegraph global warming is a fraud for world government”

​

One more Step before feeding data into model

NA in existence.

Duplicates after cleaning the data.

Watson sometimes is Paradoxical.

Majority Vote to fix the above 2 problems.

Final dataset scale: 3263

​

Methodology: Lightside

• Once we have cleansed data, we used lightside to extract features, build and evaluate models
• Data contained two columns: text/review and sentiment (Y/N)
• Very straightforward, outputted results in to excel to further inspect
• Looked at false negatives, false positives

Evaluation of Results

• We looked at accuracy as the primary metric for evaluating features
• Keeping in mind that 70% of instances are positive (belief in climate change)
• Look at kappa as well
• it doesn’t make sense to care too much about false positive or false negatives in this context
• Precision and recall relegated

​

Evaluation of Results

• Used threshold of 3, original cleaned dataset

​

Evaluation of Results: With Tags

• Used threshold of 3, original cleaned dataset

​

Without WITH

Evaluation of Results: Best Combination

Unigram+Bigram+trigram+Length

​

Discussion of Results: Best Combination

Easy classifications:

• “yeah about that global warming footprint al you hypocritical jackass”
• “some global warming huh you were all lied to by al gore he duped you”

Harder classifications:

• “wsj editorial today argues for fighting real natural disasters like tsunamis rather than man made ones like global warming” - false negative
• “what global warming #3wordslibshate” - false positive
• “tulips in mid april in chicago love the global warming” - false negative

​

Even a Harder one:

• “vry inrstg dc ok already we believe in climate change snow conts 2 pile up what s the reason” -- ?TRUE POSITIVE

​

Discussion of Results: Length

• Shorter tweets introduce more variance with features, lowering accuracy
• “tulips in mid april in chicago love the global warming” - false negative
• “what global warming #3wordslibshate” - false positive
• Longer tweets are a little bit easier:
• “larry brilliant at tedxvolcano we have to fight for science the single most important thing we face is climate change socmedia environment”

​

​

Discussion of Results: Best Combination

• Informational pieces tended to be misclassified as false negatives at higher rates
• “ultimate ebook store shocking information about global warming”
• “wsj editorial today argues for fighting real natural disasters like tsunamis rather than man made ones like global warming”
• “world famous places endangered by global warming if climatologists predictions of global warming are right some”
• “utah house passes resolution implying climate change conspiracy solveclimate com”
• Tough to tell what people are actually saying about climate change (e.g. sarcasm)
• “the only climate change that exists is summer fall winter spring just saying”

Conclusion

1. Certain topics/slang are commonly found in certain classes, seems to be corroborated by our topic analysis
1. Al Gore is a common topic amongst non-believers
2. Short reviews increase variability, decrease accuracy
3. Interestingly, unigrams were outperformed by a combination of trigrams, bigrams, and unigrams
• Context of word definitely matters more in this domain, fewer unigrams that had high information

Further research based on how views have changed over time

Questions?