def predict(self, utterance, context=list()):
# ori_word_seq = unidecode(utterance).split()
ori_word_seq = [token.text for token in self.nlp(unidecode(utterance)) if token.text.strip()]
ori_tag_seq = ['O'] * len(ori_word_seq)
if self.use_context:
if len(context) > 0 and type(context[0]) is list and len(context[0]) > 1:
context = [item[1] for item in context]
context_seq = self.dataloader.tokenizer.encode('[CLS] ' + ' [SEP] '.join(context[-3:]))
context_seq = context_seq[:512]
else:
context_seq = self.dataloader.tokenizer.encode('[CLS]')
intents = []
da = {}
word_seq, tag_seq, new2ori = self.dataloader.bert_tokenize(ori_word_seq, ori_tag_seq)
word_seq = word_seq[:512]
tag_seq = tag_seq[:512]
batch_data = [[ori_word_seq, ori_tag_seq, intents, da, context_seq,
new2ori, word_seq, self.dataloader.seq_tag2id(tag_seq), self.dataloader.seq_intent2id(intents)]]
pad_batch = self.dataloader.pad_batch(batch_data)
pad_batch = tuple(t.to(self.model.device) for t in pad_batch)
word_seq_tensor, tag_seq_tensor, intent_tensor, word_mask_tensor, tag_mask_tensor, context_seq_tensor, context_mask_tensor = pad_batch
slot_logits, intent_logits = self.model.forward(word_seq_tensor, word_mask_tensor,
context_seq_tensor=context_seq_tensor,
context_mask_tensor=context_mask_tensor)
das = recover_intent(self.dataloader, intent_logits[0], slot_logits[0], tag_mask_tensor[0],
batch_data[0][0], batch_data[0][-4])
dialog_act = []
for intent, slot, value in das:
domain, intent = intent.split('-')
dialog_act.append([intent, domain, slot, value])
return dialog_act
batch_size, data_key=data_key):
pad_batch = tuple(t.to(DEVICE) for t in pad_batch)
word_seq_tensor, tag_seq_tensor, intent_tensor, word_mask_tensor, tag_mask_tensor, context_seq_tensor, context_mask_tensor = pad_batch
if not config['model']['context']:
context_seq_tensor, context_mask_tensor = None, None
with torch.no_grad():
slot_logits, intent_logits, batch_slot_loss, batch_intent_loss = model.forward(
word_seq_tensor, word_mask_tensor, tag_seq_tensor,
tag_mask_tensor, intent_tensor, context_seq_tensor,
context_mask_tensor)
slot_loss += batch_slot_loss.item() * real_batch_size
intent_loss += batch_intent_loss.item() * real_batch_size
for j in range(real_batch_size):
predicts = recover_intent(dataloader, intent_logits[j],
slot_logits[j], tag_mask_tensor[j],
ori_batch[j][0], ori_batch[j][-4])
labels = ori_batch[j][3]
predict_golden['overall'].append({
'predict': predicts,
'golden': labels
})
predict_golden['slot'].append({
'predict': [x for x in predicts if is_slot_da(x)],
'golden': [x for x in labels if is_slot_da(x)]
})
predict_golden['intent'].append({
'predict': [x for x in predicts if not is_slot_da(x)],
'golden': [x for x in labels if not is_slot_da(x)]
})