Datasets:

DFKI-SLT

/

fabner

like 1

Follow

Speech and Language Technology, DFKI 47

Hello!

Pull Request overview

• Remove 'O-MATE', 'O-MANP', etc. from the fabner subset.
• Increment version to 1.2.0

Details

Before

Currently, labels like 'O-MATE' get created, which obviously don't make much sense. I analyzed the dataset and determined how frequently each of the labels were used:

from datasets import load_dataset
import matplotlib.pyplot as plt
from collections import Counter

dataset = load_dataset("DFKI-SLT/fabner", "fabner")
labels = dataset["train"].features["ner_tags"].feature.names
num_labels = len(labels)
print(f"A total of {num_labels} labels found.")

label_counter = Counter(sum(dataset["train"]["ner_tags"], []))
print(label_counter)

unused_labels = set()
for i in range(num_labels):
    if label_counter[i] == 0:
        unused_labels.add(labels[i])

print(f"These labels are not used whatsoever: {unused_labels}")

fig, ax = plt.subplots()
ax.set_yscale("log")
ax.bar(list(label_counter.keys()), list(label_counter.values()))
plt.show()

Which outputs:

A total of 61 labels found.
Counter({0: 184499, 50: 12349, 5: 7539, 10: 4374, 46: 3085, 49: 3085, 6: 3043, 9: 3043, 30: 2272, 1: 1516, 4: 1516, 26: 1464, 29: 1464, 7: 1459, 40: 1454, 15: 1329, 20: 1200, 36: 1122, 39: 1122, 35: 955, 25: 928, 31: 783, 34: 783, 45: 679, 11: 432, 14: 432, 21: 428, 24: 428, 47: 345, 2: 257, 32: 193, 41: 179, 44: 179, 16: 156, 19: 156, 27: 96, 37: 87, 60: 75, 12: 64, 42: 47, 55: 32, 17: 32, 22: 22, 56: 16, 59: 16, 51: 6, 54: 6, 52: 1})
These labels are not used whatsoever: {'O-PRO', 'O-PARA', 'O-CHAR', 'O-CONPRI', 'O-MACEQ', 'O-MANP', 'I-BIOP', 'O-APPL', 'O-MATE', 'O-BIOP', 'O-ENAT', 'O-MANS', 'O-FEAT'}

as well as:

Certainly not ideal! A lot of "O-..." labels that we don't want. The problem is this line, which unintuitively adds "O-..." labels:
https://huggingface.co./datasets/DFKI-SLT/fabner/blob/main/fabner.py#L136

After

After removing this line (and updating the version), running this same script gives:

A total of 49 labels found.
Counter({0: 184499, 40: 12349, 4: 7539, 8: 4374, 37: 3085, 39: 3085, 5: 3043, 7: 3043, 24: 2272, 1: 1516, 3: 1516, 21: 1464, 23: 1464, 6: 1459, 32: 1454, 12: 1329, 16: 1200, 29: 1122, 31: 1122, 28: 955, 20: 928, 25: 783, 27: 783, 36: 679, 9: 432, 11: 432, 17: 428, 19: 428, 38: 345, 2: 257, 26: 193, 33: 179, 35: 179, 13: 156, 15: 156, 22: 96, 30: 87, 48: 75, 10: 64, 34: 47, 44: 32, 14: 32, 18: 22, 45: 16, 47: 16, 41: 6, 43: 6, 42: 1})
These labels are not used whatsoever: {'I-BIOP'}

as well as:

As you can see, I-BIOP is not represented in this dataset, but that is fine. After all, in the BIOES tagging scheme, the I-... is only used for entities of 3 or more words.

Let me know if you need anything from me at this point.

• Tom Aarsen