Day 4 — Proposal Drafts
Day 4 is where we turn the verified literature into a concrete research plan: sharper research questions, testable hypotheses, and an experimental design that fits the programme’s constraints.
Daily Recap
Before we design any experiments, we should check whether our original problem statement still holds up after a week of reading — the literature might have answered our intial RQs, or present a different direction that could be more exciting! In either case, thinking about pivoting here is a sign of good research instincts, not failure. Good researchers update their beliefs as evidence comes in, and getting comfortable doing that is a big part of learning to take risks.
We started with “can we detect health misinformation in Singlish?” Day 3’s reading turned up several strong detection models that already handle code-mixed text — but none of them report whether their explanations stay faithful when the input is code-mixed. So we pivot from building a detector to evaluating explanation faithfulness on code-mixed input. Narrower, less crowded, and more clearly a contribution.
Designing Experiments
Recall that a research question (RQ) says what I want to find out to address the problem statement; a hypothesis says what I expect the answer to be, in a form I can test empirically. The RQs I’m happiest with are specific, grounded in a gap from the literature, and answerable within my constraints.
Again, this is just one way to form RQs/hypothesis. In other domains of work, proving theoretical bounds for instance, the process of forming these RQs could vary! Work together with your mentors/advisors to find the method most suitable for your flavour of research.
| Weak | Strong | |
|---|---|---|
| RQ | “How can AI detect misinformation?” | “Do token-attribution explanations remain faithful when health-misinformation classifiers are applied to code-mixed English–Singlish text?” |
| Why | Too broad, not answerable, no gap | Specific, tied to a documented gap, testable in 5 weeks |
Then I pair each RQ with a testable hypothesis:
RQ: Do token-attribution explanations remain faithful for code-mixed Singlish input?
Hypothesis: Explanation faithfulness (measured by attribution–accuracy agreement under token deletion) is significantly lower for code-mixed inputs than for monolingual English inputs, because the model attends disproportionately to English tokens.
What makes this testable, to me: it names a metric, a comparison, and a directional prediction with a stated mechanism.
As a team, draft 1–3 research questions for your (possibly pivoted) problem. For each, write a testable hypothesis that names what you’d measure and what you expect. Then poke holes in each other’s formulations: is it specific? grounded in a gap? answerable in the time you have?
Proposal Outlines
A good experimental plan in a proposal, to me, shows how I’d answer each RQ and stays honest about feasibility. In a short sprint, I lean toward plans that reuse existing datasets and models rather than ones that need new data collection or training from scratch.
- Data: an existing English health-misinformation dataset + a small set of code-mixed examples (translated/adapted, or sourced from public Singlish corpora within PDPA limits).
- Models: two off-the-shelf pre-trained classifiers with attribution methods (no training from scratch).
- Measure: faithfulness via token-deletion agreement, compared across English vs. code-mixed inputs.
- Expected outcome: a quantified faithfulness gap, plus qualitative examples of where attributions break down.
- Limitations: small code-mixed sample; results indicative rather than definitive.
- Budget: last but certainly not the least, would be to keep a rough estimate of the costs for this project (e.g., API costs, GPU hours, compute wall-clock etc.)
The proposal should minimally contain these points:
- Problem — the gap and why it matters (from Day 1, refined).
- Literature / Introduction — what exists and what it leaves open (from Days 2–3).
- Proposed direction — the RQs, hypotheses, and experimental plan.
- Expected outcomes — what a meaningful result looks like, and the limitations.
Sketch a one-page experimental plan for your main RQ (data, method, measure, expected outcome, limitations), then arrange your material into the four-part arc above. Keep it skeletal for now — the full prose is tomorrow’s job.
Team Review
Each team gives a brief walkthrough of its proposal outline and gathers peer feedback to refine before drafting.
A proposal outline containing the research question(s), paired hypotheses, and an experimental plan — structured along the four-part arc. This is the scaffold you’ll write up on Day 5.