class DialogueManager(object):
def __init__(self, agent, model):
self.agent = agent
self.dialogue_stack = DialogueStack(agent, agent.memory)
self.model = model
def get_safety_words(self, safety_words_path):
"""Read a list of safety words to prevent abuse."""
with open(safety_words_path) as f:
safety_lines = f.readlines()
safety_words = []
for l in safety_lines:
w = l.strip("\n").lower()
if w != "" and w[0] != "<" and w[0] != "#":
safety_words.append(w)
return safety_words
def is_safe(self, string):
safety_set = set(self.safety_words)
cmd_set = set(string.lower().split())
notsafe = len(cmd_set & safety_set) > 0
return not notsafe
# the dialogue manager model should access the task stack and chat history
# through the agent's memory, adding the most recent chat here as convenience
# maybe add a get_new_chat to memory and don't pass?
# chat is a (speaker, str) tuple
def step(self, chat: Tuple[str, str]):
# check safety
if not self.is_safe(chat[1]):
self.dialogue_stack.append_new(Say, "Please don't be rude.")
return
if chat[1]:
logging.info("Dialogue stack pre-run_model: {}".format(
self.dialogue_stack.stack))
# NOTE: the model is responsible for not putting a new
# object on the stack if it sees that whatever is on
# the stack should continue.
# TODO: Maybe we need a HoldOn dialogue object?
obj = self.maybe_get_dialogue_obj(chat)
if obj is not None:
self.dialogue_stack.append(obj)
# Always call dialogue_stack.step(), even if chat is empty
if len(self.dialogue_stack) > 0:
self.dialogue_stack.step()
def maybe_get_dialogue_obj(
self, chat: Tuple[str, str]) -> Optional[DialogueObject]:
raise NotImplementedError(
"Must implement maybe_get_dialogue_object in subclass")
class DialogueManager(object):
"""
| The current control flow of dialogue is:
| 1. A chat comes in and Dialogue manager reads it or the bot triggers a
| dialogue because of memory/perception/task state
| 2. The dialogue manager puts a DialogueObject on the DialogueStack.
| 3. The DialogueStack calls .step() which in turn calls the DialogueObject.step()
| that performs some action as implemented in the step method. The step could
| also possibly interact with the agent's memory. And finally the step() makes
| a call and decides if the DialogueObject is finished.
|
|
| - The step() returns a string: maybe_chat, a dict: maybe_data.
| - The step()'s outputs are read by the manager which can decide to put another
| DialogueObject on the stack.
| The maybe_data from the output of the dialogue object's step() can
| contain a 'push' key; this overrides the manager's decision on what to push to
| the stack.
| Control flow for interpreter and clarification:
| The interpreter is also a kind of DialogueObject, and a clarification step is
| the interpreter returning control to the DialogueManager, which pushes a
| ConfirmTask or ConfirmReferenceObject as a DialogueObject onto the DialogueStack.
args:
agent: a droidlet agent
model: a (perhaps ML) model used by and the model used for manager.
"""
def __init__(self, agent, model):
self.agent = agent
self.dialogue_stack = DialogueStack(agent, agent.memory)
self.model = model
def get_safety_words(self, safety_words_path):
"""Read a list of safety words to prevent abuse.
"""
with open(safety_words_path) as f:
safety_lines = f.readlines()
safety_words = []
for l in safety_lines:
w = l.strip("\n").lower()
if w != "" and w[0] != "<" and w[0] != "#":
safety_words.append(w)
return safety_words
def is_safe(self, chat):
"""Check that chat does not contain unsafe words.
"""
safety_set = set(self.safety_words)
cmd_set = set(chat.lower().split())
notsafe = len(cmd_set & safety_set) > 0
return not notsafe
def step(self, chat: Tuple[str, str]):
"""Process a chat and step through the dialogue manager task stack.
The chat is given as input to the model, which returns a logical form.
The logical form is converted to a dialogue object, which allows the interpreter to
handle the action.
Unless empty, the dialogue object is put on the dialogue stack.
Then the DialogueStack calls .step(), which in turn calls the DialogueObject.step().
DialogueObject.step() determines whether the dialogue object has finished.
Args:
chat (Tuple[str, str]): Tuple of (speaker, chat). Speaker is name of speaker.
Chat is an incoming chat command that the agent has received.
Example: ("player_1", "build a red cube")
"""
# check safety
if not self.is_safe(chat[1]):
self.dialogue_stack.append_new(Say, "Please don't be rude.")
return
if chat[1]:
logging.debug("Dialogue stack pre-run_model: {}".format(
self.dialogue_stack.stack))
# NOTE: the model is responsible for not putting a new
# object on the stack if it sees that whatever is on
# the stack should continue.
# TODO: Maybe we need a HoldOn dialogue object?
obj = self.maybe_get_dialogue_obj(chat)
if obj is not None:
self.dialogue_stack.append(obj)
# Always call dialogue_stack.step(), even if chat is empty
if len(self.dialogue_stack) > 0:
self.dialogue_stack.step()
def maybe_get_dialogue_obj(
self, chat: Tuple[str, str]) -> Optional[DialogueObject]:
raise NotImplementedError(
"Must implement maybe_get_dialogue_object in subclass")