Research Methodology

Research Methodology

A practical, end-to-end guide for doing rigorous CS/AI/ML/NLP research — from first literature search to a submittable paper. Built for someone who already codes and runs experiments (e.g. Kaggle/NLP work) but wants a repeatable, structured research process.

The core loop

Research is rarely linear, but a project usually cycles through these four phases. Each has its own reference file with full detail — read the one that matches where the user is right now.

1. Literature review → references/literature-review.md Find what's already been done, spot the gap your work fills.
2. Research design → references/research-design.md Turn the gap into a testable question, design the dataset/experiment, pick metrics and baselines.
3. Academic writing → references/academic-writing.md Structure and write the paper/thesis, handle citations, target the right venue.
4. Tools & workflow → references/tools-workflow.md Reference managers, Overleaf/LaTeX, experiment tracking, Git — the plumbing that keeps a project reproducible.

How to use this skill

• If the user names a phase directly ("help me find related work", "how do I write the methodology section"), jump straight to that reference file.
• If they're starting fresh, walk them through the loop in order: help scope a research question first (phase 2 lightly), then go back and do a proper literature review, then flesh out the full design, then write.
• If they already have a draft (paper, dataset, or code) and want feedback, diagnose which phase is weakest before jumping to advice — a weak literature review often looks like a weak methodology section later.
• Always keep advice concrete and actionable: point to a specific search query, a specific metric, a specific section heading — not just abstract principles.

Quick-start checklist for a new project

Use this as a first pass with the user, then dive into the relevant reference file for depth on any step:

1. Write a one-sentence research question. If it can't be answered "yes/no" or with a number, tighten it.
2. Search 3 sources minimum (e.g. Google Scholar, arXiv, ACL Anthology for NLP) for the last 3–5 years of related work.
3. Build a literature matrix (see assets/lit-review-tracker-template.csv) — don't just save PDFs, extract what each paper actually contributes.
4. Identify the gap: what does no existing paper do that this project will?
5. Decide the evaluation: what metric proves the work succeeded, and what's the baseline it must beat?
6. Set up version control and a reference manager before writing a single paragraph (see references/tools-workflow.md).
7. Draft the paper outline early (see assets/paper-outline-template.md) — writing the shape of the paper before all results are in exposes weak sections sooner.

Notes for CS/AI/ML/NLP specifically

• A "literature review" for ML work often means checking leaderboards (Papers With Code) and recent arXiv preprints, not just journals — coverage moves faster than formal publication.
• For Bangla/low-resource NLP work specifically, also check ACL Anthology and LREC proceedings, which index most Bangla NLP papers that Google Scholar under-indexes.
• Dataset-driven papers (like MRC or classification datasets) need their own rigor: documented annotation guidelines, inter-annotator agreement, and a clear train/val/test split with no leakage — see references/research-design.md.
• If the goal is a Scopus-indexed or Q1 journal, venue choice affects paper structure and length — check references/academic-writing.md before writing the first draft, not after.