title: "TLUnified-NER Corpus"
description: |

  - **Homepage:** [Github](https://github.com/ljvmiranda921/calamanCy)
  - **Repository:** [Github](https://github.com/ljvmiranda921/calamanCy)
  - **Point of Contact:** ljvmiranda@gmail.com

  ### Dataset Summary

  This dataset contains the annotated TLUnified corpora from Cruz and Cheng
  (2021).  It is a curated sample of around 7,000 documents for the
  named entity recognition (NER) task.  The majority of the corpus are news
  reports in Tagalog, resembling the domain of the original ConLL 2003.  There
  are three entity types: Person (PER), Organization (ORG), and Location (LOC).

  | Dataset     | Examples | PER  | ORG  | LOC  |
  |-------------|----------|------|------|------|
  | Train       | 6252     | 6418 | 3121 | 3296 |
  | Development | 782      | 793  | 392  | 409  |
  | Test        | 782      | 818  | 423  | 438  |

  ### Data Fields

  The data fields are the same among all splits:
  - `id`: a `string` feature
  - `tokens`: a `list` of `string` features.
  - `ner_tags`: a `list` of classification labels, with possible values including `O` (0), `B-PER` (1), `I-PER` (2), `B-ORG` (3), `I-ORG` (4), `B-LOC` (5), `I-LOC` (6)

  ### Annotation process

  The author, together with two more annotators, labeled curated portions of
  TLUnified in the course of four months. All annotators are native speakers of
  Tagalog.  For each annotation round, the annotators resolved disagreements,
  updated the annotation guidelines, and corrected past annotations. They
  followed the process prescribed by [Reiters
  (2017)](https://nilsreiter.de/blog/2017/howto-annotation).

  They also measured the inter-annotator agreement (IAA) by computing pairwise
  comparisons and averaging the results:
  - Cohen's Kappa (all tokens): 0.81
  - Cohen's Kappa (annotated tokens only): 0.65
  - F1-score: 0.91

  ### About this repository

  This repository is a [spaCy project](https://spacy.io/usage/projects) for
  converting the annotated spaCy files into IOB. The process goes like this: we
  download the raw corpus from Google Cloud Storage (GCS), convert the spaCy
  files into a readable IOB format, and parse that using our loading script
  (i.e., `tlunified-ner.py`). We're also shipping the IOB file so that it's
  easier to access.

directories: ["assets", "corpus/spacy", "corpus/iob"]

vars:
  version: 1.0

assets:
  - dest: assets/corpus.tar.gz
    description: "Annotated TLUnified corpora in spaCy format with train, dev, and test splits."
    url: "https://storage.googleapis.com/ljvmiranda/calamanCy/tl_tlunified_gold/v${vars.version}/corpus.tar.gz"

workflows:
  all:
    - "setup-data"
    - "upload-to-hf"

commands:
  - name: "setup-data"
    help: "Prepare the Tagalog corpora used for training various spaCy components"
    script:
      - mkdir -p corpus/spacy
      - tar -xzvf assets/corpus.tar.gz -C corpus/spacy
      - python -m spacy_to_iob corpus/spacy/ corpus/iob/
    outputs:
      - corpus/iob/train.iob
      - corpus/iob/dev.iob
      - corpus/iob/test.iob

  - name: "upload-to-hf"
    help: "Upload dataset to HuggingFace Hub"
    script:
      - git push
    deps:
      - corpus/iob/train.iob
      - corpus/iob/dev.iob
      - corpus/iob/test.iob