mycroft-core/mycroft/skills/intent_service.py

518 lines
20 KiB
Python

# Copyright 2017 Mycroft AI Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import time
from adapt.context import ContextManagerFrame
from adapt.engine import IntentDeterminationEngine
from adapt.intent import IntentBuilder
from mycroft.configuration import Configuration
from mycroft.messagebus.message import Message
from mycroft.util.lang import set_active_lang
from mycroft.util.log import LOG
from mycroft.util.parse import normalize
from mycroft.metrics import report_timing, Stopwatch
from mycroft.skills.padatious_service import PadatiousService
from .intent_service_interface import open_intent_envelope
class AdaptIntent(IntentBuilder):
def __init__(self, name=''):
super().__init__(name)
def workaround_one_of_context(best_intent):
""" Handle Adapt issue with context injection combined with one_of.
For all entries in the intent result where the value is None try to
populate using a value from the __tags__ structure.
"""
for key in best_intent:
if best_intent[key] is None:
for t in best_intent['__tags__']:
if key in t:
best_intent[key] = t[key][0]['entities'][0]['key']
return best_intent
class ContextManager:
"""
ContextManager
Use to track context throughout the course of a conversational session.
How to manage a session's lifecycle is not captured here.
"""
def __init__(self, timeout):
self.frame_stack = []
self.timeout = timeout * 60 # minutes to seconds
def clear_context(self):
self.frame_stack = []
def remove_context(self, context_id):
self.frame_stack = [(f, t) for (f, t) in self.frame_stack
if context_id in f.entities[0].get('data', [])]
def inject_context(self, entity, metadata=None):
"""
Args:
entity(object): Format example...
{'data': 'Entity tag as <str>',
'key': 'entity proper name as <str>',
'confidence': <float>'
}
metadata(object): dict, arbitrary metadata about entity injected
"""
metadata = metadata or {}
try:
if len(self.frame_stack) > 0:
top_frame = self.frame_stack[0]
else:
top_frame = None
if top_frame and top_frame[0].metadata_matches(metadata):
top_frame[0].merge_context(entity, metadata)
else:
frame = ContextManagerFrame(entities=[entity],
metadata=metadata.copy())
self.frame_stack.insert(0, (frame, time.time()))
except (IndexError, KeyError):
pass
def get_context(self, max_frames=None, missing_entities=None):
""" Constructs a list of entities from the context.
Args:
max_frames(int): maximum number of frames to look back
missing_entities(list of str): a list or set of tag names,
as strings
Returns:
list: a list of entities
"""
missing_entities = missing_entities or []
relevant_frames = [frame[0] for frame in self.frame_stack if
time.time() - frame[1] < self.timeout]
if not max_frames or max_frames > len(relevant_frames):
max_frames = len(relevant_frames)
missing_entities = list(missing_entities)
context = []
last = ''
depth = 0
for i in range(max_frames):
frame_entities = [entity.copy() for entity in
relevant_frames[i].entities]
for entity in frame_entities:
entity['confidence'] = entity.get('confidence', 1.0) \
/ (2.0 + depth)
context += frame_entities
# Update depth
if entity['origin'] != last or entity['origin'] == '':
depth += 1
last = entity['origin']
print(depth)
result = []
if len(missing_entities) > 0:
for entity in context:
if entity.get('data') in missing_entities:
result.append(entity)
# NOTE: this implies that we will only ever get one
# of an entity kind from context, unless specified
# multiple times in missing_entities. Cannot get
# an arbitrary number of an entity kind.
missing_entities.remove(entity.get('data'))
else:
result = context
# Only use the latest instance of each keyword
stripped = []
processed = []
for f in result:
keyword = f['data'][0][1]
if keyword not in processed:
stripped.append(f)
processed.append(keyword)
result = stripped
return result
class IntentService:
def __init__(self, bus):
self.config = Configuration.get().get('context', {})
self.engine = IntentDeterminationEngine()
# Dictionary for translating a skill id to a name
self.skill_names = {}
# Context related intializations
self.context_keywords = self.config.get('keywords', [])
self.context_max_frames = self.config.get('max_frames', 3)
self.context_timeout = self.config.get('timeout', 2)
self.context_greedy = self.config.get('greedy', False)
self.context_manager = ContextManager(self.context_timeout)
self.bus = bus
self.bus.on('register_vocab', self.handle_register_vocab)
self.bus.on('register_intent', self.handle_register_intent)
self.bus.on('recognizer_loop:utterance', self.handle_utterance)
self.bus.on('detach_intent', self.handle_detach_intent)
self.bus.on('detach_skill', self.handle_detach_skill)
# Context related handlers
self.bus.on('add_context', self.handle_add_context)
self.bus.on('remove_context', self.handle_remove_context)
self.bus.on('clear_context', self.handle_clear_context)
# Converse method
self.bus.on('skill.converse.response', self.handle_converse_response)
self.bus.on('skill.converse.error', self.handle_converse_error)
self.bus.on('mycroft.speech.recognition.unknown', self.reset_converse)
self.bus.on('mycroft.skills.loaded', self.update_skill_name_dict)
def add_active_skill_handler(message):
self.add_active_skill(message.data['skill_id'])
self.bus.on('active_skill_request', add_active_skill_handler)
self.active_skills = [] # [skill_id , timestamp]
self.converse_timeout = 5 # minutes to prune active_skills
self.waiting_for_converse = False
self.converse_result = False
self.converse_skill_id = ""
def update_skill_name_dict(self, message):
"""
Messagebus handler, updates dictionary of if to skill name
conversions.
"""
self.skill_names[message.data['id']] = message.data['name']
def get_skill_name(self, skill_id):
""" Get skill name from skill ID.
Args:
skill_id: a skill id as encoded in Intent handlers.
Returns:
(str) Skill name or the skill id if the skill wasn't found
"""
return self.skill_names.get(skill_id, skill_id)
def reset_converse(self, message):
"""Let skills know there was a problem with speech recognition"""
lang = message.data.get('lang', "en-us")
set_active_lang(lang)
for skill in self.active_skills:
self.do_converse(None, skill[0], lang)
def do_converse(self, utterances, skill_id, lang):
self.waiting_for_converse = True
self.converse_result = False
self.converse_skill_id = skill_id
self.bus.emit(Message("skill.converse.request", {
"skill_id": skill_id, "utterances": utterances, "lang": lang}))
start_time = time.time()
t = 0
while self.waiting_for_converse and t < 5:
t = time.time() - start_time
time.sleep(0.1)
self.waiting_for_converse = False
self.converse_skill_id = ""
return self.converse_result
def handle_converse_error(self, message):
skill_id = message.data["skill_id"]
if message.data["error"] == "skill id does not exist":
self.remove_active_skill(skill_id)
if skill_id == self.converse_skill_id:
self.converse_result = False
self.waiting_for_converse = False
def handle_converse_response(self, message):
skill_id = message.data["skill_id"]
if skill_id == self.converse_skill_id:
self.converse_result = message.data.get("result", False)
self.waiting_for_converse = False
def remove_active_skill(self, skill_id):
for skill in self.active_skills:
if skill[0] == skill_id:
self.active_skills.remove(skill)
def add_active_skill(self, skill_id):
# search the list for an existing entry that already contains it
# and remove that reference
self.remove_active_skill(skill_id)
# add skill with timestamp to start of skill_list
self.active_skills.insert(0, [skill_id, time.time()])
def update_context(self, intent):
""" Updates context with keyword from the intent.
NOTE: This method currently won't handle one_of intent keywords
since it's not using quite the same format as other intent
keywords. This is under investigation in adapt, PR pending.
Args:
intent: Intent to scan for keywords
"""
for tag in intent['__tags__']:
if 'entities' not in tag:
continue
context_entity = tag['entities'][0]
if self.context_greedy:
self.context_manager.inject_context(context_entity)
elif context_entity['data'][0][1] in self.context_keywords:
self.context_manager.inject_context(context_entity)
def send_metrics(self, intent, context, stopwatch):
"""
Send timing metrics to the backend.
NOTE: This only applies to those with Opt In.
"""
ident = context['ident'] if 'ident' in context else None
if intent:
# Recreate skill name from skill id
parts = intent.get('intent_type', '').split(':')
intent_type = self.get_skill_name(parts[0])
if len(parts) > 1:
intent_type = ':'.join([intent_type] + parts[1:])
report_timing(ident, 'intent_service', stopwatch,
{'intent_type': intent_type})
else:
report_timing(ident, 'intent_service', stopwatch,
{'intent_type': 'intent_failure'})
def handle_utterance(self, message):
""" Main entrypoint for handling user utterances with Mycroft skills
Monitor the messagebus for 'recognizer_loop:utterance', typically
generated by a spoken interaction but potentially also from a CLI
or other method of injecting a 'user utterance' into the system.
Utterances then work through this sequence to be handled:
1) Active skills attempt to handle using converse()
2) Padatious high match intents (conf > 0.95)
3) Adapt intent handlers
5) Fallbacks:
- Padatious near match intents (conf > 0.8)
- General fallbacks
- Padatious loose match intents (conf > 0.5)
- Unknown intent handler
Args:
message (Message): The messagebus data
"""
try:
# Get language of the utterance
lang = message.data.get('lang', "en-us")
set_active_lang(lang)
utterances = message.data.get('utterances', [])
# normalize() changes "it's a boy" to "it is a boy", etc.
norm_utterances = [normalize(u.lower(), remove_articles=False)
for u in utterances]
# Build list with raw utterance(s) first, then optionally a
# normalized version following.
combined = utterances + list(set(norm_utterances) -
set(utterances))
LOG.debug("Utterances: {}".format(combined))
stopwatch = Stopwatch()
intent = None
padatious_intent = None
with stopwatch:
# Give active skills an opportunity to handle the utterance
converse = self._converse(combined, lang)
if not converse:
# No conversation, use intent system to handle utterance
intent = self._adapt_intent_match(utterances,
norm_utterances, lang)
for utt in combined:
_intent = PadatiousService.instance.calc_intent(utt)
if _intent:
best = padatious_intent.conf if padatious_intent\
else 0.0
if best < _intent.conf:
padatious_intent = _intent
LOG.debug("Padatious intent: {}".format(padatious_intent))
LOG.debug(" Adapt intent: {}".format(intent))
if converse:
# Report that converse handled the intent and return
LOG.debug("Handled in converse()")
ident = message.context['ident'] if message.context else None
report_timing(ident, 'intent_service', stopwatch,
{'intent_type': 'converse'})
return
elif (intent and intent.get('confidence', 0.0) > 0.0 and
not (padatious_intent and padatious_intent.conf >= 0.95)):
# Send the message to the Adapt intent's handler unless
# Padatious is REALLY sure it was directed at it instead.
self.update_context(intent)
# update active skills
skill_id = intent['intent_type'].split(":")[0]
self.add_active_skill(skill_id)
# Adapt doesn't handle context injection for one_of keywords
# correctly. Workaround this issue if possible.
try:
intent = workaround_one_of_context(intent)
except LookupError:
LOG.error('Error during workaround_one_of_context')
reply = message.reply(intent.get('intent_type'), intent)
else:
# Allow fallback system to handle utterance
# NOTE: A matched padatious_intent is handled this way, too
# TODO: Need to redefine intent_failure when STT can return
# multiple hypothesis -- i.e. len(utterances) > 1
reply = message.reply('intent_failure',
{'utterance': utterances[0],
'norm_utt': norm_utterances[0],
'lang': lang})
self.bus.emit(reply)
self.send_metrics(intent, message.context, stopwatch)
except Exception as e:
LOG.exception(e)
def _converse(self, utterances, lang):
""" Give active skills a chance at the utterance
Args:
utterances (list): list of utterances
lang (string): 4 letter ISO language code
Returns:
bool: True if converse handled it, False if no skill processes it
"""
# check for conversation time-out
self.active_skills = [skill for skill in self.active_skills
if time.time() - skill[
1] <= self.converse_timeout * 60]
# check if any skill wants to handle utterance
for skill in self.active_skills:
if self.do_converse(utterances, skill[0], lang):
# update timestamp, or there will be a timeout where
# intent stops conversing whether its being used or not
self.add_active_skill(skill[0])
return True
return False
def _adapt_intent_match(self, raw_utt, norm_utt, lang):
""" Run the Adapt engine to search for an matching intent
Args:
raw_utt (list): list of utterances
norm_utt (list): same list of utterances, normalized
lang (string): language code, e.g "en-us"
Returns:
Intent structure, or None if no match was found.
"""
best_intent = None
def take_best(intent, utt):
nonlocal best_intent
best = best_intent.get('confidence', 0.0) if best_intent else 0.0
conf = intent.get('confidence', 0.0)
if conf > best:
best_intent = intent
# TODO - Shouldn't Adapt do this?
best_intent['utterance'] = utt
for idx, utt in enumerate(raw_utt):
try:
intents = [i for i in self.engine.determine_intent(
utt, 100,
include_tags=True,
context_manager=self.context_manager)]
if intents:
take_best(intents[0], utt)
# Also test the normalized version, but set the utternace to
# the raw version so skill has access to original STT
norm_intents = [i for i in self.engine.determine_intent(
norm_utt[idx], 100,
include_tags=True,
context_manager=self.context_manager)]
if norm_intents:
take_best(norm_intents[0], utt)
except Exception as e:
LOG.exception(e)
return best_intent
def handle_register_vocab(self, message):
start_concept = message.data.get('start')
end_concept = message.data.get('end')
regex_str = message.data.get('regex')
alias_of = message.data.get('alias_of')
if regex_str:
self.engine.register_regex_entity(regex_str)
else:
self.engine.register_entity(
start_concept, end_concept, alias_of=alias_of)
def handle_register_intent(self, message):
intent = open_intent_envelope(message)
self.engine.register_intent_parser(intent)
def handle_detach_intent(self, message):
intent_name = message.data.get('intent_name')
new_parsers = [
p for p in self.engine.intent_parsers if p.name != intent_name]
self.engine.intent_parsers = new_parsers
def handle_detach_skill(self, message):
skill_id = message.data.get('skill_id')
new_parsers = [
p for p in self.engine.intent_parsers if
not p.name.startswith(skill_id)]
self.engine.intent_parsers = new_parsers
def handle_add_context(self, message):
""" Add context
Args:
message: data contains the 'context' item to add
optionally can include 'word' to be injected as
an alias for the context item.
"""
entity = {'confidence': 1.0}
context = message.data.get('context')
word = message.data.get('word') or ''
origin = message.data.get('origin') or ''
# if not a string type try creating a string from it
if not isinstance(word, str):
word = str(word)
entity['data'] = [(word, context)]
entity['match'] = word
entity['key'] = word
entity['origin'] = origin
self.context_manager.inject_context(entity)
def handle_remove_context(self, message):
""" Remove specific context
Args:
message: data contains the 'context' item to remove
"""
context = message.data.get('context')
if context:
self.context_manager.remove_context(context)
def handle_clear_context(self, message):
""" Clears all keywords from context """
self.context_manager.clear_context()