convert_czner.py

import os
import random

from datasets import load_dataset


def whitespace_tokenize_with_offsets(text):
    tokens = []
    start_tok_offsets = []
    end_tok_offsets = []
    current_token = ""
    current_token_start = None

    for i, char in enumerate(text):
        if char.isspace():
            if current_token:
                tokens.append(current_token)
                start_tok_offsets.append(current_token_start)
                end_tok_offsets.append(i)
                current_token = ""
                current_token_start = None
        else:
            if current_token == "":
                current_token_start = i
            current_token += char

    # Add the last token if there is one
    if current_token:
        tokens.append(current_token)
        start_tok_offsets.append(current_token_start)
        end_tok_offsets.append(len(text))

    return tokens, start_tok_offsets, end_tok_offsets


def proc_dataset(dataset, max_text_length=200):
    r = []
    for doc in dataset:
        text = doc["text"]
        covered_entities = set()
        for ent_id, entity in enumerate(doc["entities"]):
            if ent_id in covered_entities:
                continue
            target_text = text
            if len(text) > max_text_length:
                tokens, start_tok_offsets, end_tok_offsets = whitespace_tokenize_with_offsets(text)
                entity_start = entity["start"]
                entity_end = entity["end"]

                # Find the token indices that correspond to the entity
                entity_start_idx = None
                entity_end_idx = None
                for idx, (start, end) in enumerate(zip(start_tok_offsets, end_tok_offsets)):
                    if start <= entity_start < end:
                        entity_start_idx = idx
                    if start < entity_end <= end:
                        entity_end_idx = idx
                        break

                if entity_start_idx is None or entity_end_idx is None:
                    continue

                allowed_tokens = max_text_length - len(tokens[entity_start_idx:entity_end_idx + 1]) - 20
                before_tokens = random.randint(0, int(allowed_tokens * 0.8))
                after_tokens = allowed_tokens - before_tokens

                # Determine the start and end indices for the new text segment
                if entity_start_idx - before_tokens < 0:
                    after_tokens += - (entity_start_idx - before_tokens)
                elif entity_end_idx + after_tokens + 1 >= len(tokens):
                    before_tokens += entity_end_idx + after_tokens + 1 - len(tokens)
                start_idx = max(0, entity_start_idx - before_tokens)
                end_idx = min(len(tokens), entity_end_idx + after_tokens + 1)

                # Ensure the first 20 tokens are included if possible
                initial_text = ""
                if start_idx > 20:
                    initial_text = text[:end_tok_offsets[20]] + "... "

                # Use offsets to extract the original text
                start_offset = start_tok_offsets[start_idx]
                end_offset = end_tok_offsets[end_idx - 1]

                target_text = initial_text + text[start_offset:end_offset]

            # if target text contains more entities of the same type, add them to the answers and covered entities
            this_answer_entities = [ent_id]
            answers = [entity["content"]]
            for ent_id2, entity2 in enumerate(doc["entities"]):
                if ent_id2 == ent_id:
                    continue
                # check type
                if entity2["category_str"] == entity["category_str"]:
                    # just check the string in the target text
                    # check if the entity is in the target text
                    if entity2["content"] in target_text:
                        this_answer_entities.append(ent_id2)
                        answers.append(entity2["content"])

            covered_entities.update(this_answer_entities)

            r.append({
                "label": entity["category_str"],
                "answers": list(set(answers)),
                "text": target_text,
            })
    return r


d = load_dataset("fewshot-goes-multilingual/cs_czech-named-entity-corpus_2.0")
train = list(d['train'])
random.shuffle(train)
new_dataset_train = proc_dataset(train[3000:])
dataset_test_ftrain = proc_dataset(train[:3000])
dataset_val = proc_dataset(d['validation'])
dataset_test = proc_dataset(d['test'])

# merge splits
new_dataset_test = dataset_test_ftrain + dataset_val + dataset_test
random.shuffle(new_dataset_test)

# save using jsonlines in .data/hf_datasets/ner_court_decisions
os.makedirs(".data/hf_dataset/czner_2.0", exist_ok=True)
import jsonlines

# print dataset lengths
print("train", len(new_dataset_train))
print("test", len(new_dataset_test))

with jsonlines.open(".data/hf_dataset/czner_2.0/train.jsonl", "w") as f:
    f.write_all(new_dataset_train)
with jsonlines.open(".data/hf_dataset/czner_2.0/test.jsonl", "w") as f:
    f.write_all(new_dataset_test)