def get_pos_matrix(self, pos_list):
pos_emb = [
self.pos_embeddings(
utils.set_cuda(torch.tensor(word_pos), self.cuda_on))
for word_pos in pos_list
]
return torch.stack(pos_emb)
def get_ner_matrix(self, ner_list):
ner_emb = [
self.ner_embeddings(
utils.set_cuda(torch.tensor(word_ner), self.cuda_on))
for word_ner in ner_list
]
return torch.stack(ner_emb)
def compute_similarity(self, word, question):
similarity = []
word_value = word.lower() if word.lower(
) in self.words_embeddings.vocab else 'unk'
for question_word in question:
question_value = question_word.lower() if question_word.lower(
) in self.words_embeddings.vocab else 'unk'
similarity.append(
self.words_embeddings.similarity(word_value, question_value))
return utils.set_cuda(torch.tensor(sum(similarity)), self.cuda_on)
def forward(self, paragraph, question):
paragraph_emb = self.get_sentence_embeddings(paragraph['context'])
paragraph_pos = paragraph['context_pos']
paragraph_ner = paragraph['context_ner']
paragraph_match = torch.stack([
utils.set_cuda(torch.tensor(match), self.cuda_on)
for match in question['exact_match']
])
question_emb = self.get_sentence_embeddings(question['question'])
paragraph_vector = self.create_doc_vector(paragraph_emb, paragraph_pos,
paragraph_ner,
paragraph_match,
paragraph['context'],
question['question'])
size = question_emb.size()
question_vector = self.qFFN(question_emb.view(1, size[0], size[1]))
size = paragraph_vector.size()
paragraph_vector = self.dFFN(paragraph_vector.view(
1, size[0], size[1]))
return paragraph_vector, question_vector, paragraph_emb, question_emb
def get_word_embeddings(self, word):
return utils.set_cuda(
torch.tensor(
self.words_embeddings[word.lower()] if word.lower() in
self.words_embeddings.vocab else self.words_embeddings['unk']),
self.cuda_on)