Mental-State Verbs and Theory of Mind

ISEF Category: Behavioral and Social Sciences

Subcategory: Development  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

Kids can say words like think, know, and want long before they fully explain what another person believes. That gap makes a great research target. The CHILDES corpus gives you transcript data from real conversations, so you can check when those verbs first appear and compare that timing with theory-of-mind milestones. You get a clean question with a big idea behind it, does language help kids learn to think about minds, or do the two move together?

What Is It?

Mental-state verbs are words people use to talk about thoughts, beliefs, wishes, and knowledge. Think, know, and want are the classic examples. In child language data, these words matter because they may mark when a child starts talking about inner states, not just objects or actions. This project treats the CHILDES corpus like a giant timeline of everyday speech. You can look for the first meaningful use of each verb in each language, then compare that timing with published theory-of-mind milestones, such as false-belief understanding. Picture it like checking whether children hear the labels for mental states before they can solve the puzzle of another person's point of view.

Why This Is a Good Topic

This is a strong science fair topic because the question is specific, measurable, and tied to a real developmental debate. You can work with public transcripts and published milestone data, so you do not need a lab bench. You also get to practice corpus search, coding rules, and basic stats, which are useful skills for any later psychology, linguistics, or data project.

Research Questions

• How does the age of first meaningful use of think, know, or want in CHILDES relate to later theory-of-mind milestone ages within the same language?
• What is the effect of language family on the gap between first mental-state verb use and first false-belief success?
• Does the amount of caregiver mental-state speech in early transcripts predict the child's age at first mental-state verb use?
• To what extent does bilingual exposure change the gap between first mental-state verb use and theory-of-mind milestone age?
• Which verb, think, know, or want, gives the strongest match to later theory-of-mind timing across corpora?
• How does transcript density change the relationship between first mental-state verb use and milestone age?

Basic Materials

• Laptop with internet access.
• Access to CHILDES transcripts on TalkBank.
• Spreadsheet software such as Google Sheets or Excel.
• A text editor for cleaning transcript files.
• A published table of theory-of-mind milestone ages.

Advanced Materials

• Python environment with Jupyter Notebook.
• R and RStudio for statistical modeling.
• pandas for cleaning transcript metadata and age tables.
• spaCy or NLTK for transcript parsing and verb search.
• A coding notebook or annotation sheet for reliability checks.

Software & Tools

• Python: Cleans transcript text, counts target verbs, and merges corpus data with milestone tables.
• Jupyter Notebook: Keeps code, notes, and output together while you test the analysis.
• pandas: Organizes transcript metadata and builds age-by-verb summaries.
• R: Fits regression or mixed-effects models for cross-language comparisons.
• spaCy: Finds verb forms and nearby context in large text files.

Experiment Steps

1. Define one mental-state verb set and one theory-of-mind milestone set before you touch the data.
2. Choose the languages, age bands, and transcript filters that will keep your sample comparable.
3. Set a coding rule for first meaningful use, then decide how you will check borderline cases.
4. Build a table that links corpus ages to milestone ages, and decide how you will normalize for transcript count.
5. Pick the statistical test or model that matches your data, then plan one sensitivity check for tiny samples or outlier corpora.

Common Pitfalls

• Counting adult speech as child speech, which makes the verb seem earlier than it is.
• Treating formulaic phrases like I do not know as proof of full mental-state understanding, which overstates the signal.
• Mixing corpora with very different sizes, which makes rich datasets look stronger than sparse ones.
• Comparing milestone studies that use different tasks, which turns one outcome into several different outcomes.
• Skipping normalization for transcript density, which blurs whether you found a development pattern or just more words.

What Makes This Competitive

A stronger project goes beyond a simple correlation. You would predefine what counts as first use, control for corpus size and language family, and test whether the link still holds after you remove tiny samples. A stronger entry could compare child speech and caregiver speech, or use time-to-event analysis to ask when each verb first appears relative to each milestone. That kind of design shows you understand both development and data.

Project Variations

• Compare bilingual and monolingual CHILDES corpora to see whether mental-state verbs appear at different ages.
• Swap child speech for caregiver speech, and test whether adult mental-state input predicts later milestone timing.
• Compare false-belief milestones with appearance-reality milestones to see whether the pattern depends on the task.

Learn More

• TalkBank CHILDES: Browse child language transcripts, corpus documentation, and search tools on the TalkBank site.
• PubMed: Search review articles on mental-state language, theory of mind, and language development.
• NCBI Bookshelf: Read free background chapters on child development, cognition, and language acquisition.
• MIT OpenCourseWare: Use free linguistics or statistics courses for analysis background.
• Open Science Framework: Find public preregistrations and shared analysis templates for developmental studies.