Study Notes

MNLP - Mini Project

5 min read

MNLP Mini Project

Source

Based on Christof Monz’s Mini Project slides (PDF in Assets/). 9-slide Beamer presentation covering scope, timeline, deliverables, and quality criteria.

Overview

The MNLP mini project is a six-week group project (Weeks 2-7 of Period 1) where teams of ~5 students work on one concrete multilingual NLP problem. It’s the main practical component of the course alongside the exam.

Each team picks a problem (from the Canvas list or self-designed), implements neural approaches, compares architectures, does error analysis, and presents findings in a 4-page report + 10-15 minute presentation.

Schedule & Deadlines

DateWhat
Fri 2026-09-04, 17:00Submit team + problem choice via Google Sheet (link on Canvas)
Week 2-6Project work (see weekly breakdown below)
Fri 2026-10-09, 12:00Submit report, slides, and code
Week 7Team presentations (10-15 min + 5 min Q&A)

Weekly Breakdown

Week 1 (Now)

  • Choose your team (5 students)
  • Pick a problem from the Canvas list, or design your own
  • If self-designed: email Monz a 1-page description

Week 2 — Literature & Planning

  • Identify and read relevant previous literature
  • Inspect data, pre-processing, split into train/val/test
  • Additional data crawling if needed
  • Decide on an initial (baseline) model

Week 3 — First Model

  • Implement the first model
  • Evaluate
  • Debug
  • Document: what works, what doesn’t?

Week 4 — Refinement

  • Refine first model or try an alternative model
  • Bells and whistles: dropout, layer norms, more layers, residual connections, highway connections, etc.
  • Evaluate and compare
  • Document: what works, what doesn’t?

Week 5 — Error Analysis

  • Refine second model / try another alternative model
  • Error analysis, open problems
  • Conclusions, recommendations

Week 6 — Wrap Up

  • Address any loose ends
  • Write the 4-page report with main findings
  • Prepare presentation slides

Week 7 — Presentations

  • Presentation slots assigned at beginning of Week 6
  • 10-15 minutes presentation + 5 minutes for questions

Project Requirements

Team Formation

  • Teams of 5 students
  • Fill out the Google Sheet by Sep 4, 17:00 with:
    • All team members (full names, student IDs, emails)
    • The problem you choose to work on
    • Meeting time slot preferences (no guarantees)
  • If designing your own problem: email Monz a 1-page description

Problem Selection

  • Canvas course page lists example problems with links to relevant datasets
  • You can also choose your own problem (subject to approval via 1-page description)

Deliverables (due Oct 9, 12:00 noon)

DeliverableDetails
Report~4 pages, PDF format; focus on most relevant findings
SlidesPDF or PowerPoint
CodeGitHub repo link (preferred) or gzipped tarball; must include readme explaining how to run
DataIf gzipped data < 10 MB: include it; otherwise provide a link

Evaluation Criteria

See the “What Makes a Good Project” section below. The evaluation focuses on:

  • Quality of literature review and motivation
  • Soundness of preprocessing and architecture choices
  • Thoroughness of evaluation (multiple settings/architectures + error analysis)
  • Clarity of report and presentation

Practical Advice

Keep Data Small

Most training will run on your laptop. If data is too big:

  • Down-sample
  • Select based on criteria: vocabulary, classes, length, etc.

What Makes a Good Mini Project

Monz’s slides list six quality dimensions:

  1. Literature Review — good, succinct description of the most relevant research papers
  2. Data Description — good description of preprocessing steps and experimental settings
  3. Architecture Motivation — good motivation for why you chose the neural architectures you did
  4. Thorough Evaluation — scores under different settings/architectures, plus error analysis
  5. Focused Report — the report should center on the most relevant findings: what works and what doesn’t, and why
  6. Strong Presentation — focus on the most important aspects and findings, understandable by a wider audience, and include examples

Example: How to “Pimp Up” a Project (Language Identification)

Monz provides a detailed walkthrough using language identification as a case study, showing how to elevate a seemingly simple problem into a thorough project:

Literature & Input

  • What has been tried? (neural and non-neural approaches)
  • Process input: words, characters, or character -grams?

Models

  1. Baseline (Week 2): word embeddings — how to combine embeddings; additional hidden layers?
  2. Better model (Week 3): CNN sequence classifier — how to pool; strides, etc.
  3. Even better model (Week 3/4): Transformer sequence classifier

Evaluation Dimensions

  • Standard accuracies (check literature for benchmarks)
  • Short input problem (literature mentions this)
  • Languages close to each other (e.g., closely related language families)
  • Code-switching cases
  • How accuracy changes after characters (useful for instant translation applications)
  • Best overall metrics: accuracy, precision, recall, F1, etc.

Data & Features

  • Crawl/use more data
  • Script information: can UTF-8 region information be used in a neural network?

Presentation of Findings

  • Score tables
  • Visualizations
  • Clusters
  • Error analysis: what really makes a difference?
  • Anything that contradicts previous literature? Is it a bug or a genuine finding?

Key Takeaways

The Core Loop

Literature Baseline Refine Evaluate Analyze Errors Conclude

  • 6 weeks of work, culminating in a 4-page report, slides, and code due Oct 9 at noon
  • Teams of 5 — form your team and pick a problem by Sep 4, 17:00
  • The bar isn’t just “does it work” — it’s “why does (or doesn’t) it work?” with thorough error analysis
  • Keep data manageable for laptop training
  • Presentations in Week 7 (10-15 min + 5 min Q&A)

Graph View

Backlinks