Climate Misinformation in Regional News Comments

ISEF Category: Earth and Environmental Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Climate misinformation does not spread evenly. A county that deals with floods may talk very differently from one that faces drought, even when both read the same news. You can test that pattern with text data, not guesses.

What Is It?

This project studies how people talk about climate policy online. You collect public comments on climate-related news stories, then look for scientific misconceptions, such as claims that a warming climate cannot affect local weather. In simple terms, you are teaching a computer to sort comments by idea, the way a librarian sorts books by topic.

The machine learning part usually uses a transformer, which is a type of language model that learns patterns in text. You can fine-tune a pre-trained model, which means you start with a model that already understands language and then train it on your own labeled examples. Your goal is not just to count negative comments. You want to see whether certain misconceptions appear more often in places with specific climate risks, such as flood-prone or drought-prone counties.

Why This Is a Good Topic

This topic works well for a science fair because it gives you a clear question, measurable data, and a real-world problem. You can connect online misinformation to public understanding of climate risk, which matters for communication, policy, and education. You will also learn skills that researchers use in NLP, data labeling, model testing, and statistical comparison. The project feels current, but it still has enough structure for a student to finish with careful planning.

Research Questions

• How does county-level flood risk relate to the frequency of specific climate misconceptions in public comments?
• How does county-level drought risk relate to the frequency of specific climate misconceptions in public comments?
• Does the rate of climate misinformation differ between comments on policy articles and comments on science explainer articles?
• To what extent do misinformation labels change when you compare a keyword filter with a fine-tuned transformer?
• Which misconception categories appear most often in comments from high-impact versus low-impact counties?
• What is the effect of article topic, such as heat, floods, or emissions, on the type of misinformation that appears?

Basic Materials

• Computer with internet access and enough storage for text files.
• Spreadsheet software such as Google Sheets or Excel.
• Python installed through Anaconda or a similar free distribution.
• Jupyter Notebook for cleaning text and testing models.
• Public climate risk data from NOAA, USGS, or county government sources.
• Publicly available news comments or archived discussion data that allow research use.
• A codebook for labeling misinformation categories.
• A sample set of manually labeled comments for training and evaluation.

Advanced Materials

• Access to a university server or cloud notebook with GPU support.
• Python environment with PyTorch, Hugging Face Transformers, pandas, scikit-learn, and spaCy.
• Annotation tool such as Label Studio for building labeled datasets.
• Geographic data files for counties, hazard zones, and census boundaries.
• Statistical software or Python packages for regression and mixed-effects modeling.
• Data management plan for storing raw text, labels, and model outputs separately.
• Ethics review or advisor support for handling public-facing text data.
• Version control with Git for code and analysis tracking.

Software & Tools

• Python: Cleans text, trains classifiers, and runs statistical summaries.
• Jupyter Notebook: Keeps code, notes, and results in one place.
• Hugging Face Transformers: Lets you fine-tune a pre-trained language model on labeled comments.
• pandas: Organizes comment data, labels, and county-level climate variables.
• scikit-learn: Measures classifier performance and compares baseline models.

Experiment Steps

1. Define the misinformation categories you will track and write clear labeling rules.
2. Choose a public comment source and a matching county-level climate impact dataset.
3. Build a labeled sample so you can train and test your text model.
4. Select a baseline approach and a transformer model so you can compare performance.
5. Plan the county-level comparison, including how you will group regions by climate risk.
6. Decide which statistics will test whether misinformation patterns differ across regions.

Common Pitfalls

• Using too few labeled comments, which makes the model learn noise instead of real patterns.
• Mixing county-level climate data with user-level comment data without a clear join rule, which breaks the regional comparison.
• Treating all negative or skeptical comments as misinformation, which inflates the label.
• Training and testing on nearly identical comments, which makes accuracy look better than it is.
• Ignoring class imbalance, which can hide rare but important misconception categories.

What Makes This Competitive

A strong version of this project goes beyond a simple count of false claims. You can compare a baseline keyword method with a fine-tuned model, then test whether the model finds patterns that the simple method misses. You can also separate misinformation by claim type, not just by sentiment. Careful validation, clear regional grouping, and a smart statistical test will make the project much stronger.

Project Variations

• Study comments on wildfire news instead of flood and drought news, then compare misinformation patterns across fire-risk regions.
• Classify misinformation by stance, such as denial, delay, or false solution claims, instead of by broad topic.
• Compare comments from national news outlets with comments from local news outlets to see whether geography changes the pattern.

Learn More

• PubMed: Search for review articles on climate misinformation, risk perception, and health communication.
• NOAA Climate.gov: Find background on regional climate impacts and plain-language explanations of hazards.
• USGS National Water Information System: Use stream, flood, and drought data for regional comparisons.
• NASA Earthdata: Explore climate and weather datasets that help define regional impact variables.
• MIT OpenCourseWare, Introduction to Machine Learning: Review free course materials on classification, model evaluation, and overfitting.
• Nature Climate Change: Search for peer-reviewed studies on climate communication and misinformation.