README.md

metadata

dataset_info:
  features:
    - name: id
      dtype: int32
    - name: tokens
      list: string
    - name: entities
      list: string
    - name: ner_tags
      list:
        class_label:
          names:
            '0': O
            '1': B-Person-Individual
            '2': I-Person-Individual
            '3': B-Person-Collective
            '4': I-Person-Collective
            '5': B-Organization-Political
            '6': I-Organization-Political
            '7': B-Organization-Government
            '8': I-Organization-Government
            '9': B-Organization-Military
            '10': I-Organization-Military
            '11': B-Organization-Other
            '12': I-Organization-Other
            '13': B-Location
            '14': I-Location
            '15': B-Time
            '16': I-Time
            '17': B-Production-Media
            '18': I-Production-Media
            '19': B-Production-Government
            '20': I-Production-Government
            '21': B-Production-Doctrine
            '22': I-Production-Doctrine
            '23': B-Numerical Statistics
            '24': I-Numerical Statistics
            '25': B-Object-Weapon
            '26': I-Object-Weapon
            '27': B-Event
            '28': I-Event
    - name: Row_Index
      dtype: int32
    - name: Year
      dtype: string
    - name: Publication
      dtype: string
  splits:
    - name: train
      num_bytes: 6294856
      num_examples: 5602
    - name: validation
      num_bytes: 892624
      num_examples: 799
    - name: test
      num_bytes: 1865217
      num_examples: 1599
  download_size: 1376214
  dataset_size: 9052697
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
task_categories:
  - token-classification
language:
  - tl
tags:
  - history
  - philippines
  - martial-law
pretty_name: ph-martial_law-ner
size_categories:
  - 1K<n<10K

Dataset Summary

This dataset is sourced from various Martial Law newspaper repositories, namely University of Hawaii’s eVols and Bantayog ng mga Bayani digital library. In total, the dataset contains approximately ~13k sentences and segmented into 8000 samples. The table below displays the entities—formulated by following the annotation guidelines—and its token distribution.

Dataset	Examples	PER-IND	PER-COLL	ORG-POL	ORG-GOV	ORG-MIL	ORG-OTHER	LOC	TIME	PROD-MEDIA	PROD-GOV	PROD-DOCT	STAT	EVENT	OBJ-WPN
Train	5602	3359	876	893	1460	958	918	3118	1396	538	518	1336	2144	896	332
Development	799	487	132	144	208	137	129	492	199	77	78	195	238	159	50
Test	1599	981	265	235	383	330	242	994	440	161	127	426	656	257	104

Data Fields

• id : an int assigned id in the sample's respective dataset
• tokens : a list tokenized, delimited by whitespace
• entities : a list of classification labels, in IOB (Inside, Outside, Beginning) format
• ner_tags : a list of numeric values, 0 for outside, 1 to 15 maps to an entity
• Row_Index : an int indicating original row index of sample in raw data spreadsheet
• Year : a string year of publication
• Publication : a string, publisher

Annotation Process

Annotating the dataset involved three annotators and was performed in an iterative fashion. After each batch, the researchers and annotators would convene to discuss and clarify any confusion regarding the annotation guidelines, which were then revised. In spite of these efforts, the Inter-annotator agreement (IAA) did not meet the standard threshold for a reliable annotation score (0.8). As such, this final batch

• Cohen's Kappa (all tokens): 0.86
• Cohen's Kappa (annotated tokens only): 0.72
• F1-score: 0.74 Hence, The dataset went through various quality assurance stages, ultimately leading to the description of the dataset presented here. Full guidelines can be found in this Google Doc.

TODO:

• Trained models (and performers)