README.md

metadata

dataset_info:
  features:
    - name: chars
      sequence: string
    - name: labels
      sequence:
        class_label:
          names:
            '0': D
            '1': I
            '2': P
            '3': S
  splits:
    - name: train
      num_bytes: 2298515
      num_examples: 12290
    - name: validation
      num_bytes: 37643
      num_examples: 221
    - name: test
      num_bytes: 44203
      num_examples: 257
  download_size: 341126
  dataset_size: 2380361
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
language:
  - yue
license: cc-by-4.0
task_categories:
  - token-classification

This data is the subset of the Hong Kong Cantonese Corpus (HKCanCor) that has been re-segmented by the multi-tiered word segmentation scheme described in the following paper:

Charles Lam, Chaak-ming Lau, and Jackson L. Lee. 2024. Multi-Tiered Cantonese Word Segmentation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11993–12002, Torino, Italy. ELRA and ICCL.

Processing from original format

Chinese word segmentation is commonly framed as a sequence labelling task. To ease the adoption of the original dataset, we have transformed it into two columns:

1. chars: The original sentence split into characters
2. labels: One of D (Dash), I (Intermediate), P (Pipe), S (Space). This labelling scheme expands the common BI (Beginning/Intermediate) scheme into four labels. In particular, the I is the same as the BI scheme but the B tag is further split into D, P, and S.

Here's a sample segmented sentence in the original format and the processed format:

Original: 即係 噉樣 嗰-啲 呀 ?

Processed:

• chars: 即 係 噉 樣 嗰 啲 呀 ?
• labels: S I S I S D S S

Note how the label is taken from the left boundary of the labelled character. For the left most character in a sentence, there is no boundary character to its left. In these cases, we set up a convention to always label the left most character as S (Space).

Check out https://github.com/AlienKevin/dips for more details.