Yvthyvq/CantoMap-Liujgoj

# CantoMap-Liujgoj

This repository contains a comprehensively transliterated and streamlined version of the CantoMap corpus, strictly converted into **Liujgoj (溜歌粵語)**—a standalone, character-free orthography designed for computational linguistics and Cantonese AI architecture.

The original data belongs to the gwinterstein/CantoMap project. This derivative version strips away heavy unlinked binary assets to provide a lightweight, high-density textual corpus tailored for Cantonese natural language processing.

---

🎯 Key Modifications

1. Strict Liujgoj Transliteration: Every standard Jyutping syllable has been mapped onto the Liujgoj orthographic inventory, replacing numerical tone marks with letter-based tone markers (j, r, x, q, h).
2. Purged Large File Storage (LFS) Overhead: All broken or unlinked binary/multimedia placeholders (.wav audio shells and .pdf maps) have been completely removed. The repository is 100% text-driven, lightweight (~1.65 MiB), and optimized for instant cloning/streaming.

---

📦 Dataset Features

• Format & Structure: Contains 102 EAF (ELAN Annotation Format) files across organized subfolders, maintaining the exact conversational structure of the original Map Task.
• Preserved Linguistic Integrity:
  • Word Boundaries: Retains the exact word-segmentation and token spacing intended by the original transcribers.
  • Colloquial Markers: Preserves all native speech-act annotations, filled pauses (#), and lengthening markers.
  • Variant Pronunciations: Safely preserves phonetic variant tags (e.g., waih|wair, gam|gamr), making it a goldmine for analyzing morphological tone changes and spontaneous phonetic drifts.

---

🤖 AI & LLM Training Applications

This corpus is highly valuable for developing Speech-Native Multimodal AI and training Cantonese Large Language Models (LLMs):

• Stage 1 Continual Pre-training (CPT): The pure alphabetical stream of Liujgoj eliminates tokenization fragmentation caused by mixed alphanumeric strings (like laa1aa1).
• Turn-taking & Interactivity: The interactive Map Task structure provides raw data for fine-tuning conversational agents on genuine, collaborative human dialogue.

Quick Python Snippet to Parse Text

You can easily extract the Liujgoj text tiers for language modeling using standard XML parsing:

import xml.etree.ElementTree as ET

def extract_liujgoj(eaf_path):
    tree = ET.parse(eaf_path)
    root = tree.getroot()
    texts = []
    for tier in root.findall(".//TIER"):
        tier_id = tier.get("TIER_ID", "")
        if tier_id.endswith("-jyutping") or tier_id == "jyutping":
            for ann in tier.findall(".//ANNOTATION_VALUE"):
                if ann.text:
                    texts.append(ann.text)
    return texts

---

🎵 關於音頻檔案 / Audio Files

⚠️ 注意 / Note: 本數據集資料夾內原本的音頻檔案已損壞。 如需獲取本數據集對應的完整音頻，請前往以下網址下載：

👉 safecantonese/cantomap

下載音頻後，請依據音頻檔名與本數據集內的文本/標記檔案（如 CSV 等）進行對齊使用。

---

📄 Attribution & License Attribution The original data and speech-act annotations belong to the CantoMap project created by gwinterstein and collaborators.

Original Repository: gwinterstein/CantoMap

If you use this dataset, please ensure you credit and cite the original parent project accordingly.

License This derivative dataset is distributed under the CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International) license, strictly inheriting the original licensing terms of the parent CantoMap corpus.