Awesome

ner_dataset_recognition

Dataset information

The datasets are provided in the datasets file.

• The once with _sentences contain the sentences and the conference information
• The once without it are in IOB2 format

Datasets explained

Only information about the datasets without _sentences are provided, _sentences contain the same sentences as there counterpart

• Dataset.csv - The whole dataset

• Zero-shot.csv - Contains 200 sentences with datasets that do not occur in the train or test set

• Train_set.csv - The 80% of the 80/20 split of the remaining sentences of the whole dataset (so without the zero-shot set)

• Test_set.csv - The 20% of the 80/20 split of the remaining sentences of the whole dataset (so without the zero-shot set)

• Test_set_real_ratio - A test set where only 1% of the sentences contain dataset names

• Test_set_easy - 444 sentences from the test set that only contain dataset names with three or fewer words

• Test_set_hard - 444 sentences from the test set that only contain dataset names with four or more words

• SSC - A dataset containing only weak supervised examples

• SSC_positives - Only the potive weak supervised examples

Code information

The code is providid in the code file

• The files with _cross-validation are the code for cross-validation
• The other ones are the regular methods.

Getting the right data

In the code the standard situation is shown, so training on the train set and testing on the test set.

For the more specialised cases, code is provided below

Domain adaptability

DATAset = pd.read_csv('Dataset_sentences.csv')

BIOset = pd.read_csv('Dataset.csv')

ids = DATAset[DATAset.conference == 'VISION'].id.to_list()

sentences = ['Sentence: ' + str(id) for id in ids]

data = BIOset[~BIOset['Sentence #'].isin(sentences)]

test = BIOset[BIOset['Sentence #'].isin(sentences)]

Amount of training data

This is done via slicing, slices of the stratisfied 20 fold are used and each time added to the previous slice

Settings:

skf = StratifiedKFold(20, shuffle=True, random_state=42)

Positive/negative ratio

The right data is selected using slicing, example: np.array(X_tr)[:2168]

TRAINset = pd.read_csv('Train_set_sentences.csv')

ids = TRAINset[~TRAINset.labels.str.contains('Geen')].id.to_list()

sentences = ['Sentence: ' + str(id) for id in ids]

ds = data[data['Sentence #'].isin(sentences)]

nds = data[~data['Sentence #'].isin(sentences)]

data = ds.append(nds)

Other

This is done by appending, example:

data = pd.read_csv('Train_set.csv') data2 = pd.read_csv('SSC.csv') data = data.append(data2)