def get_internal_state(self, cand_kb, last_turn):
agent_act, user_value = last_turn
# slot cls and inform cls
agent_act_id = AgentActs.slot_to_id(agent_act)
turn_slot_cls = torch.zeros(AgentActs.slot_size())
turn_slot_cls[agent_act_id] = 1
# user have no answer
if user_value is None:
turn_inform_cls = 1
turn_level_doc_dist = [1 for _ in range(len(cand_kb))]
else:
turn_inform_cls = 0
turn_level_doc_dist = []
for kb in cand_kb:
if not self.enable_full_answer:
if agent_act not in kb or user_value in kb[agent_act]:
turn_level_doc_dist.append(1)
continue
else:
if ', ' in user_value:
user_value = user_value.split(', ')
if agent_act not in kb or kb[
agent_act] == user_value or kb[agent_act] == [
user_value
]:
turn_level_doc_dist.append(1)
continue
turn_level_doc_dist.append(float('-inf'))
turn_level_doc_dist = torch.softmax(torch.tensor(turn_level_doc_dist,
dtype=torch.float),
dim=-1)
return turn_level_doc_dist, turn_slot_cls, turn_inform_cls
def __getitem__(self, index):
turn = self.turn_data[index]
tar_name = turn['tar_name']
cand_docs_names = turn['cand_names']
cand_docs_idx = [self.all_docs_names.index(name) for name in cand_docs_names]
cand_doc_tensor = self.docs_rep[cand_docs_idx, :]
# get the dialog history tensor
dia_his_tensor = self.qa_to_tensor(turn['dialog_history']).long()
# get the ground truth of agent act and value
agent_act = AgentActs.slot_to_id(turn['agent_act'])
# target documents and documents distribution
tar_idx = cand_docs_names.index(tar_name)
docs_dist = torch.tensor(turn['docs_dist'], dtype=torch.float).gt(0).long().squeeze(0)
return cand_doc_tensor, dia_his_tensor, agent_act, tar_idx, docs_dist
def run_dialog(self, cand_docs, cand_docs_diff, cand_names, tar_kb,
tar_name):
"""
Simulate a dialog with model
:param cand_docs:
:param cand_names:
:param tar_kb:
:param tar_name:
:return:
"""
self.agent.init_dialog(cand_docs, cand_docs_diff, cand_names)
self.simulator.init_dialog(tar_kb, tar_name)
tar_idx = cand_names.index(tar_name)
# dialog status
top_1_success = False
top_3_success = False
mrr = 0
all_rewards = []
saved_log_act_probs = [
] # saved action log probability for REINFORCE algorithm
saved_log_doc_probs = [
] # saved docs log probability for REINFORCE algorithm
num_turns = self.max_turns
# dialog start with user
# user_act, user_value, user_nl = self.simulator.respond_act(agent_act=None, agent_value=None)
# dialog_json = [{'turn_id': 0,
# 'agent_act': '',
# 'agent_value': '',
# 'agent_nl': '',
# 'user_act': user_act,
# 'user_value': user_value,
# 'user_nl': user_nl,
# 'tar_rank': 0,
# 'docs_entropy': 0}]
# if 'rule-' in self.agent_type:
# last_turn = (user_act, user_value)
# else:
# last_turn = self.dataset.turn_to_tensor(agent_nl='', user_nl=user_nl).unsqueeze(0)
dialog_his = '_BOS_'
dialog_json = []
last_turn = None
assert self.max_turns > 1
for turn_i in range(1, self.max_turns + 1):
agent_act, agent_act_prob, agent_value, agent_nl = self.agent.turn_act(
last_turn)
# feedback from environments
if len(dialog_json) == 0:
is_end = agent_act == AgentActs.GUESS
else:
reward, top_1_success, top_3_success, is_end, mrr, tar_r, docs_entropy, tar_prob = \
self.env_feedback(agent_act,
self.agent.dialog_level_doc_dist,
tar_idx)
all_rewards.append(reward)
dialog_json[-1]['tar_rank'] = tar_r
dialog_json[-1]['tar_prob'] = float('{:.2f}'.format(tar_prob))
dialog_json[-1]['docs_entropy'] = float(
'{:.2f}'.format(docs_entropy))
# user response
user_act, user_value, user_nl = self.simulator.respond_act(
agent_act, agent_value)
# record current turn
if 'rule-' in self.agent_type or 'mrc' in self.agent_type:
last_turn = (user_act, user_value)
else:
last_turn = self.dataset.turn_to_tensor(
agent_nl=agent_nl, user_nl=user_nl).unsqueeze(0)
# record data
turn_json = {
'turn_id': turn_i,
'agent_act': agent_act,
'agent_value': agent_value,
'agent_nl': agent_nl,
'user_act': user_act,
'user_value': user_value,
'user_nl': user_nl,
'tar_rank': 0,
'tar_prob': 0,
'docs_entropy': 0,
'docs_dist': self.agent.dialog_level_doc_dist.tolist()
}
dialog_json.append(turn_json)
# early stop
if is_end:
num_turns = turn_i
break
# saved log probability for rl-training
if self.training:
agent_act_id = AgentActs.slot_to_id(agent_act)
saved_log_act_probs.append(
torch.log(agent_act_prob[agent_act_id]))
pred_doc_id = cand_names.index(agent_value)
saved_log_doc_probs.append(
torch.log(
self.agent.dialog_level_doc_dist[0][pred_doc_id]))
# steps on episode
self.num_steps += 1
# when training
if self.training and len(all_rewards):
loss = self.agent.update(all_rewards, saved_log_act_probs,
saved_log_doc_probs) # update the model
self.on_training(loss)
cur_rewards = sum(all_rewards)
data_json = {
'tar_name': tar_name,
'cand_names': cand_names,
'dialog': dialog_json,
'top_1_success': top_1_success,
'top_3_success': top_3_success,
'mrr': mrr,
'rewards': '{:.2f}'.format(cur_rewards)
}
return data_json, top_1_success, top_3_success, mrr, num_turns, cur_rewards