Interdisciplinary Approach to Identify and Characterize COVID-19 Misinformation on Twitter: Mixed Methods Study
Bibliographic record
Abstract
BACKGROUND: Studying COVID-19 misinformation on Twitter presents methodological challenges. A computational approach can analyze large data sets, but it is limited when interpreting context. A qualitative approach allows for a deeper analysis of content, but it is labor-intensive and feasible only for smaller data sets. OBJECTIVE: We aimed to identify and characterize tweets containing COVID-19 misinformation. METHODS: Tweets geolocated to the Philippines (January 1 to March 21, 2020) containing the words coronavirus, covid, and ncov were mined using the GetOldTweets3 Python library. This primary corpus (N=12,631) was subjected to biterm topic modeling. Key informant interviews were conducted to elicit examples of COVID-19 misinformation and determine keywords. Using NVivo (QSR International) and a combination of word frequency and text search using key informant interview keywords, subcorpus A (n=5881) was constituted and manually coded to identify misinformation. Constant comparative, iterative, and consensual analyses were used to further characterize these tweets. Tweets containing key informant interview keywords were extracted from the primary corpus and processed to constitute subcorpus B (n=4634), of which 506 tweets were manually labeled as misinformation. This training set was subjected to natural language processing to identify tweets with misinformation in the primary corpus. These tweets were further manually coded to confirm labeling. RESULTS: Biterm topic modeling of the primary corpus revealed the following topics: uncertainty, lawmaker's response, safety measures, testing, loved ones, health standards, panic buying, tragedies other than COVID-19, economy, COVID-19 statistics, precautions, health measures, international issues, adherence to guidelines, and frontliners. These were categorized into 4 major topics: nature of COVID-19, contexts and consequences, people and agents of COVID-19, and COVID-19 prevention and management. Manual coding of subcorpus A identified 398 tweets with misinformation in the following formats: misleading content (n=179), satire and/or parody (n=77), false connection (n=53), conspiracy (n=47), and false context (n=42). The discursive strategies identified were humor (n=109), fear mongering (n=67), anger and disgust (n=59), political commentary (n=59), performing credibility (n=45), overpositivity (n=32), and marketing (n=27). Natural language processing identified 165 tweets with misinformation. However, a manual review showed that 69.7% (115/165) of tweets did not contain misinformation. CONCLUSIONS: An interdisciplinary approach was used to identify tweets with COVID-19 misinformation. Natural language processing mislabeled tweets, likely due to tweets written in Filipino or a combination of the Filipino and English languages. Identifying the formats and discursive strategies of tweets with misinformation required iterative, manual, and emergent coding by human coders with experiential and cultural knowledge of Twitter. An interdisciplinary team composed of experts in health, health informatics, social science, and computer science combined computational and qualitative methods to gain a better understanding of COVID-19 misinformation on Twitter.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.034 | 0.040 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.005 | 0.004 |
| Science and technology studies | 0.004 | 0.002 |
| Scholarly communication | 0.003 | 0.003 |
| Open science | 0.002 | 0.004 |
| Research integrity | 0.002 | 0.001 |
| Insufficient payload (model declined to judge) | 0.006 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".