forked from kubapok/en-ner-conll-2003
154 lines
4.7 KiB
Python
154 lines
4.7 KiB
Python
from collections import Counter
|
|
|
|
import pandas as pd
|
|
import torch
|
|
from torchtext.vocab import Vocab
|
|
|
|
|
|
class NERModel(torch.nn.Module):
|
|
def __init__(self, ):
|
|
super(NERModel, self).__init__()
|
|
self.emb = torch.nn.Embedding(23628, 200)
|
|
self.fc1 = torch.nn.Linear(600, 9)
|
|
|
|
def forward(self, x):
|
|
x = self.emb(x)
|
|
x = x.reshape(600)
|
|
x = self.fc1(x)
|
|
return x
|
|
|
|
|
|
def data_process(dt):
|
|
return [torch.tensor([vocab['<bos>']] + [vocab[token] for token in document] + [vocab['<eos>']], dtype=torch.long)
|
|
for document in dt]
|
|
|
|
|
|
def build_vocab(dataset):
|
|
counter = Counter()
|
|
for document in dataset:
|
|
counter.update(document)
|
|
return Vocab(counter, specials=['<unk>', '<pad>', '<bos>', '<eos>'])
|
|
|
|
|
|
def labels_process(dt):
|
|
return [torch.tensor([0] + document + [0], dtype=torch.long) for document in dt]
|
|
|
|
|
|
def predict(input_tokens, labels):
|
|
results = []
|
|
for i in range(len(input_tokens)):
|
|
line_results = []
|
|
for j in range(1, len(input_tokens[i]) - 1):
|
|
x = input_tokens[i][j - 1: j + 2].to(device)
|
|
predicted = ner_model(x.long())
|
|
result = torch.argmax(predicted)
|
|
label = labels[result]
|
|
line_results.append(label)
|
|
results.append(line_results)
|
|
|
|
return results
|
|
|
|
|
|
def features(data):
|
|
featurez = []
|
|
for sentence in data["tokens"]:
|
|
t_sentence = torch.tensor(())
|
|
for word in sentence:
|
|
temp = torch.tensor([word[0].isupper(), len(word)])
|
|
t_sentence = torch.cat((t_sentence, temp))
|
|
featurez.append(t_sentence)
|
|
|
|
return featurez
|
|
|
|
|
|
def merge_features(token_ids, tensors_list):
|
|
return [torch.cat((token, tensors_list[i])) for i, token in enumerate(token_ids)]
|
|
|
|
|
|
def process_output(lines):
|
|
result = []
|
|
for line in lines:
|
|
last_label = None
|
|
new_line = []
|
|
for label in line:
|
|
if label != "O" and label[0:2] == "I-":
|
|
if last_label is None or last_label == "O":
|
|
label = label.replace('I-', 'B-')
|
|
else:
|
|
label = "I-" + last_label[2:]
|
|
last_label = label
|
|
new_line.append(label)
|
|
result.append(" ".join(new_line))
|
|
return result
|
|
|
|
|
|
def infer(path_in, path_out, labels):
|
|
df = pd.read_csv(path_in, sep='\t', names=['tokens'])
|
|
df_token_ids = data_process(df["tokens"].apply(lambda x: x.split()))
|
|
df_infer = merge_features(df_token_ids, features(df))
|
|
infers = predict(df_infer, labels)
|
|
infers_processed = process_output(infers)
|
|
with open(path_out, "w") as file_out:
|
|
for inf in infers_processed:
|
|
file_out.write(inf + "\n")
|
|
|
|
|
|
labels = ['O', 'B-LOC', 'I-LOC', 'B-MISC', 'I-MISC', 'B-ORG', 'I-ORG', 'B-PER', 'I-PER']
|
|
|
|
df = pd.read_csv('train/train.tsv.xz', compression='xz', sep='\t', names=['iob', 'tokens'])
|
|
df["iob"] = df["iob"].apply(lambda x: [labels.index(y) for y in x.split()])
|
|
df["tokens"] = df["tokens"].apply(lambda x: x.split())
|
|
|
|
|
|
vocab = build_vocab(df['tokens'])
|
|
|
|
device = torch.device("cuda:0")
|
|
ner_model = NERModel().to(device)
|
|
criterion = torch.nn.CrossEntropyLoss()
|
|
optimizer = torch.optim.Adam(ner_model.parameters())
|
|
|
|
train_labels = labels_process(df['iob'])
|
|
train_tokens_ids = data_process(df['tokens'])
|
|
df_features = features(df)
|
|
train_tensors = merge_features(train_tokens_ids, df_features)
|
|
|
|
for epoch in range(5):
|
|
acc_score = 0
|
|
prec_score = 0
|
|
selected_items = 0
|
|
recall_score = 0
|
|
relevant_items = 0
|
|
items_total = 0
|
|
ner_model.train()
|
|
for i in range(len(train_labels)):
|
|
for j in range(1, len(train_labels[i]) - 1):
|
|
X = train_tensors[i][j - 1: j + 2].to(device)
|
|
Y = train_labels[i][j: j + 1].to(device)
|
|
Y_predictions = ner_model(X.long())
|
|
|
|
acc_score += int(torch.argmax(Y_predictions) == Y)
|
|
if torch.argmax(Y_predictions) != 0:
|
|
selected_items += 1
|
|
if torch.argmax(Y_predictions) != 0 and torch.argmax(Y_predictions) == Y.item():
|
|
prec_score += 1
|
|
if Y.item() != 0:
|
|
relevant_items += 1
|
|
if Y.item() != 0 and torch.argmax(Y_predictions) == Y.item():
|
|
recall_score += 1
|
|
|
|
items_total += 1
|
|
optimizer.zero_grad()
|
|
loss = criterion(Y_predictions.unsqueeze(0), Y)
|
|
loss.backward()
|
|
optimizer.step()
|
|
|
|
precision = prec_score / selected_items
|
|
recall = recall_score / relevant_items
|
|
f1_score = (2 * precision * recall) / (precision + recall)
|
|
print(f'epoch: {epoch}')
|
|
print(f'f1: {f1_score}')
|
|
print(f'acc: {acc_score / items_total}')
|
|
|
|
infer('dev-0/in.tsv', 'dev-0/out.tsv', labels=labels)
|
|
infer('test-A/in.tsv', 'test-A/out.tsv', labels=labels)
|