process/tweet_ner.py

import os
import json
from typing import List
from datasets import load_dataset

label2id = {
    "B-corporation": 0,
    "B-creative_work": 1,
    "B-event": 2,
    "B-group": 3,
    "B-location": 4,
    "B-person": 5,
    "B-product": 6,
    "I-corporation": 7,
    "I-creative_work": 8,
    "I-event": 9,
    "I-group": 10,
    "I-location": 11,
    "I-person": 12,
    "I-product": 13,
    "O": 14
}
id2label = {v: k for k, v in label2id.items()}


def decode_ner_tags(tag_sequence: List, input_sequence: List):
    """ decode ner tag sequence """

    def update_collection(_tmp_entity, _tmp_entity_type, _tmp_pos, _out):
        if len(_tmp_entity) != 0 and _tmp_entity_type is not None:
            _out.append({'type': _tmp_entity_type, 'entity': _tmp_entity, 'position': _tmp_pos})
            _tmp_entity = []
            _tmp_entity_type = None
        return _tmp_entity, _tmp_entity_type, _tmp_pos, _out

    assert len(tag_sequence) == len(input_sequence), str([len(tag_sequence), len(input_sequence)])
    out = []
    tmp_entity = []
    tmp_pos = []
    tmp_entity_type = None
    for n, (_l, _i) in enumerate(zip(tag_sequence, input_sequence)):
        _l = id2label[_l]
        if _l.startswith('B-'):
            _, _, _, out = update_collection(tmp_entity, tmp_entity_type, tmp_pos, out)
            tmp_entity_type = '-'.join(_l.split('-')[1:])
            tmp_entity = [_i]
            tmp_pos = [n]
        elif _l.startswith('I-'):
            tmp_tmp_entity_type = '-'.join(_l.split('-')[1:])
            if len(tmp_entity) == 0:
                # if 'I' not start with 'B', skip it
                tmp_entity, tmp_entity_type, tmp_pos, out = update_collection(tmp_entity, tmp_entity_type, tmp_pos, out)
            elif tmp_tmp_entity_type != tmp_entity_type:
                # if the type does not match with the B, skip
                tmp_entity, tmp_entity_type, tmp_pos, out = update_collection(tmp_entity, tmp_entity_type, tmp_pos, out)
            else:
                tmp_entity.append(_i)
                tmp_pos.append(n)
        elif _l == 'O':
            tmp_entity, tmp_entity_type, tmp_pos, out = update_collection(tmp_entity, tmp_entity_type, tmp_pos, out)
        else:
            raise ValueError('unknown tag: {}'.format(_l))
    _, _, _, out = update_collection(tmp_entity, tmp_entity_type, tmp_pos, out)
    return out

data = load_dataset("tner/tweetner7")


def process(tmp):
    tmp = [i.to_dict() for _, i in tmp.iterrows()]
    for i in tmp:
        tokens = ["@" + t.replace("{@", "").replace("@}", "").replace(" ", "_") if t.startswith("{@") else t
                  for t in i.pop('tokens')]
        entities = []
        for e in decode_ner_tags(i.pop('tags').tolist(), tokens):
            entities.append(f'{" ".join(e["entity"])} ({e["type"]})')
        i['text'] = " ".join(tokens).replace("{{USERNAME}}", "@user").replace("{{URL}}", "{URL}")
        i['condition'] = f'Entities: {", ".join(entities)}'
    return tmp


train = process(data["train_2020"].to_pandas())
train += process(data["train_2021"].to_pandas())
val = process(data["validation_2020"].to_pandas())
val += process(data["validation_2021"].to_pandas())
test = process(data["test_2021"].to_pandas())
os.makedirs("dataset/ner", exist_ok=True)
with open("dataset/ner/train.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in train]))
with open("dataset/ner/validation.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in val]))
with open("dataset/ner/test.jsonl", "w") as f:
    f.write("\n".join([json.dumps(i) for i in test]))