Pictogram Sequence Prediction

Overview

Next-token prediction over a constrained pictogram vocabulary. Given a partial sequence of pictogram tokens, the model predicts the most likely next token using a statistical n-gram ensemble (trigram → bigram → unigram fallback with Laplace smoothing and source-word boosting).

Project Structure

IMGCLEF/
├── dataset/
│   ├── train_next_picto.json   # 63,675 training samples
│   └── test_next_picto.json    # 1,986 test samples
├── main.py                     # Prediction pipeline (train + predict + save)
├── submission.json             # Final output — 1,986 entries, 10 candidates each
├── workflow.txt                # Task specification & design notes
├── requirements.txt            # Dependencies (stdlib only)
└── README.md                   # This file

How It Works

Data Format

Each sample contains:

Field	Description
id	Unique sample identifier
src	Natural-language sentence
tgt	Corresponding pictogram token sequence
pictos	Pictogram IDs (aligned with tgt)

Model

A backoff n-gram model with Laplace smoothing:

1. Trigram — Uses the last 2 tokens as context. Primary predictor.
2. Bigram — Uses the last token. Fallback when no trigram context exists.
3. Unigram — Global token frequencies. Final fallback.

Additionally, tokens appearing in the src field receive a 1.5× score boost to leverage the natural-language signal.

Pipeline

Train JSON ──► Build n-gram counts ──► For each test sample:
                                         context tokens → score vocab
                                         → rank top 10 → emit candidates
                                       ──► submission.json

Quick Start

# No external packages needed — runs on Python 3 stdlib
python main.py

Output: submission.json with 1,986 entries, each containing an id and a ranked list of 10 candidate tokens.

Submission Format

{
  "id": "common_voice_fr_h3zlwha1.mp3",
  "candidates": ["family", "of", "by", "woman", "little_girl",
                  "house", "new", "and", "chapel", "schoolhouse"]
}

Key Statistics

Metric	Value
Training samples	63,675
Test samples	1,986
Vocabulary size	~6,841
Candidates per entry	10
External deps	None

Evaluation Metrics

The competition evaluates on:

• Top-N Accuracy (N = 1, 3, 5, 10) — is the correct token in the top N?
• Semantic Similarity — cosine similarity via all-MiniLM-L6-v2 embeddings, rewarding near-synonyms.