metadata

dataset_info:
  features:
    - name: tokens
      sequence: string
    - name: ner_tags
      sequence:
        class_label:
          names:
            '0': O
            '1': B-PER
            '2': I-PER
            '3': B-ORG
            '4': I-ORG
            '5': B-LOC
            '6': I-LOC
            '7': B-TIME
            '8': I-TIME
            '9': B-TTL
            '10': I-TTL
  splits:
    - name: train
      num_bytes: 2138256
      num_examples: 3465
  download_size: 546138
  dataset_size: 2138256
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - token-classification
language:
  - am
size_categories:
  - 1K<n<10K

Amharic Named Entity Recognition Dataset

This dataset can be used to train models for Named Entity Recognition.

Dataset Source

https://github.com/uhh-lt/ethiopicmodels/blob/master/am/data/NER/train.txt

Finetuned Models

The following transformer models were finetuned using this dataset. The reported precision, recall, and f1 metrics are macro averages.

Model	Size (# params)	Precision	Recall	F1
bert-medium-amharic	40.5M	0.64	0.73	0.68
bert-small-amharic	27.8M	0.64	0.72	0.68
bert-mini-amharic	10.7M	0.60	0.67	0.64
bert-tiny-amharic	4.18M	0.50	0.59	0.54
xlm-roberta-base	279M	0.69	0.79	0.73
am-roberta	443M	0.67	0.72	0.69

Code

In this repository, you can find notebooks for finetuning each of the above models using this dataset

https://github.com/rasyosef/amharic-named-entity-recognition