Twitter Analysis of "climate change" tweets

This repository hosts code and data for our publication "Attention to climate change only temporarily diverted by COVID-19".
Authors: Tim Repke, Max Callaghan, William Lamb, Sarah Lück, Finn Müller-Hansen, Jan Minx (all @MCC_Berlin)
Keywords: Climate Change, COVID-19 Pandemic, Public Attention, Social Media, Topic Model

---

The COVID-19 pandemic disrupted peoples' daily lives and dominated the public discourse. We analyze 13.5 million tweets on climate change during 2018--2021 and show that attention to climate dropped substantially in 2020 with the onset of the pandemic. While research has helped to explain this drop in the context of issue attention theory, our analysis highlights a remarkable recovery in attention in 2021 towards pre-pandemic levels. Moreover, our large-scale, transformer-based text analysis reveals important thematic shifts during this period. In particular, we show a sustained drop in attention to activist movements and subsequently an increased focus on climate causes and climate solutions. This means, that while the climate change discourse in general recovered from the COVID-19 pandemic, activist movements such as the school protests that have mobilized millions around the globe in 2019 have measurably lost traction on Twitter.

Structure of this repository

Data

As per Twitter ToS, we cannot share the full dataset, only the Tweet IDs.

• cc_news_coverage.xlsx: Media coverage (world-wide) of climate change by Fernandez et al.
• pwid-covid-data.csv: Number of COVID-19 cases worldwide
• english_tweet_counts_daily_2006_2021-rt.csv: Our approximation of overall daily (English-language) tweet volume
• topic_annotations_and_stats.xlsx: Annotator labels for topic to theme mappings, topic keyword lists, and statistics
• share.jsonl: Tweet IDs and topic annotations; see https://zenodo.org/record/7778199

Code

The main processing of the dataset is done by the "pipeline". All other follow-up analyses and supplemental scripts are contained in that folder as well. For example, scripts to estimate the number of clusters and heuristically determine hyper-parameters.

The utils folder contains code to deploy large computations to a SLURM cluster. This might be overly specific for reproduction, but added for completeness anyway.

Figures

• pre_post: Plots and statistics to explore how themes are more prominent before or after the pandemic after normalising overall tweet volume over time
• counts: High-res figures used in the paper
• stacked_areas: Topics per supertopic over time as stacked area charts
• quarterly_overlap.png: Semantic overlap of tweets per quarter
• topics_tweets_*.png/(daily|monthly)_(tweets|users)*.png: Number of tweets per user rates (cumulative or per time interval)
• topic_users*.png: Overlap of users posting in multiple themes