def switch_to_particular_game_discussion(vars):
user_uttr = state_utils.get_last_human_utterance(vars)
user_text = user_uttr.get("text", "").lower()
prev_bot_uttr = state_utils.get_last_bot_utterance(vars)
prev_bot_text = prev_bot_uttr.get("text", "")
found_video_game_in_user_uttr = find_games_in_text(user_text)
logger.info(
f"(switch_to_particular_game_discussion)found_video_game_in_user_uttr: {found_video_game_in_user_uttr}"
)
found_video_game_in_user_uttr = bool(found_video_game_in_user_uttr)
found_video_game_in_bot_uttr = find_games_in_text(prev_bot_text)
logger.info(
f"(switch_to_particular_game_discussion)found_video_game_in_bot_uttr: {found_video_game_in_bot_uttr}"
)
found_video_game_in_bot_uttr = bool(found_video_game_in_bot_uttr)
choose_particular_game = if_choose_topic(
user_uttr, prev_bot_uttr) and found_video_game_in_user_uttr
question_answer_contains_video_game = ("?" not in user_text
and "?" in prev_bot_text
and found_video_game_in_user_uttr)
bot_asked_about_game_and_user_answered_yes = (found_video_game_in_bot_uttr
and "?" in prev_bot_text
and is_yes(user_uttr))
flag = (lets_talk_about(
vars, GAMES_WITH_AT_LEAST_1M_COPIES_SOLD_COMPILED_PATTERN)
or choose_particular_game or question_answer_contains_video_game
or bot_asked_about_game_and_user_answered_yes)
logger.info(f"switch_to_particular_game_discussion={flag}")
return flag
def lets_chat_about_movies(uttr, prev_uttr=None):
prev_uttr = {} if prev_uttr is None else prev_uttr
curr_uttr_is_about_movies = re.search(MOVIE_PATTERN,
uttr.get("text", "").lower())
lets_talk_about_movies = if_chat_about_particular_topic(
uttr, prev_uttr, compiled_pattern=MOVIE_PATTERN)
chosed_topic = if_choose_topic(uttr,
prev_uttr) and curr_uttr_is_about_movies
if (lets_talk_about_movies or chosed_topic or
("?" not in uttr.get("text", "") and "?" in prev_uttr.get("text", "")
and curr_uttr_is_about_movies)):
return True
else:
return False
def is_about_movies(uttr, prev_uttr=None):
prev_uttr = {} if prev_uttr is None else prev_uttr
curr_uttr_is_about_movies = re.search(MOVIE_PATTERN,
uttr.get("text", "").lower())
prev_uttr_last_sent = prev_uttr.get("annotations",
{}).get("sentseg",
{}).get("segments",
[""])[-1].lower()
prev_uttr_is_about_movies = re.search(MOVIE_PATTERN, prev_uttr_last_sent)
lets_talk_about_movies = if_chat_about_particular_topic(
uttr, prev_uttr, compiled_pattern=MOVIE_PATTERN)
chosed_topic = if_choose_topic(uttr,
prev_uttr) and curr_uttr_is_about_movies
if (lets_talk_about_movies or chosed_topic or curr_uttr_is_about_movies
or ("?" in prev_uttr_last_sent and prev_uttr_is_about_movies)):
return True
else:
return False
def switch_to_general_gaming_discussion(vars):
user_uttr = state_utils.get_last_human_utterance(vars)
user_text = user_uttr.get("text", "").lower()
prev_bot_uttr = state_utils.get_last_bot_utterance(vars)
prev_bot_text = prev_bot_uttr.get("text", "")
found_video_game_words_in_user_uttr = bool(
VIDEO_GAME_WORDS_COMPILED_PATTERN.search(user_text))
choose_gaming_discussion = if_choose_topic(
user_uttr, prev_bot_uttr) and found_video_game_words_in_user_uttr
question_answer_contains_video_game_words = (
"?" not in user_text and "?" in prev_bot_text
and found_video_game_words_in_user_uttr)
bot_asked_about_game_and_user_answered_yes = is_yes(
user_uttr) and is_question_about_games(prev_bot_text)
flag = (lets_talk_about(vars, VIDEO_GAME_WORDS_COMPILED_PATTERN)
or choose_gaming_discussion
or question_answer_contains_video_game_words
or bot_asked_about_game_and_user_answered_yes)
logger.info(f"switch_to_general_gaming_discussion={flag}")
return flag
async def send(self, payload: Dict, callback: Callable):
st_time = time.time()
try:
dialog = payload["payload"]["states_batch"][0]
skills_for_uttr = []
user_uttr = dialog["human_utterances"][-1]
user_uttr_text = user_uttr["text"].lower()
user_uttr_annotations = user_uttr["annotations"]
bot_uttr = dialog["bot_utterances"][-1] if len(
dialog["bot_utterances"]) else {}
bot_uttr_text_lower = bot_uttr.get("text", "").lower()
prev_active_skill = bot_uttr.get("active_skill", "")
intent_catcher_intents = get_intents(user_uttr,
probs=False,
which="intent_catcher")
high_priority_intent_detected = any([
k for k in intent_catcher_intents
if k in high_priority_intents["dff_intent_responder_skill"]
])
low_priority_intent_detected = any([
k for k in intent_catcher_intents if k in low_priority_intents
])
detected_topics = set(get_topics(user_uttr, which="all"))
is_factoid = get_factoid(user_uttr).get("is_factoid", 0.0) > 0.96
is_celebrity_mentioned = check_is_celebrity_mentioned(user_uttr)
if_choose_topic_detected = if_choose_topic(user_uttr, bot_uttr)
if_lets_chat_about_particular_topic_detected = if_chat_about_particular_topic(
user_uttr, bot_uttr)
dialog_len = len(dialog["human_utterances"])
exit_cond = "exit" in intent_catcher_intents and (
dialog_len == 1 or
(dialog_len == 2 and len(user_uttr_text.split()) > 3))
repeat_cond = ("repeat" in intent_catcher_intents
and prev_active_skill in UNPREDICTABLE_SKILLS
and re.match(r"^what.?$", user_uttr_text))
cant_do_cond = ("cant_do" in intent_catcher_intents
and "play" in user_uttr_text and any([
phrase in bot_uttr_text_lower
for phrase in GREETING_QUESTIONS_TEXTS
]))
for intent_name, condition in zip(
["exit", "repeat", "cant_do"],
[exit_cond, repeat_cond, cant_do_cond]):
if condition:
high_priority_intent_detected = False
not_detected = {"detected": 0, "confidence": 0.0}
user_uttr["annotations"]["intent_catcher"][
intent_name] = not_detected
dialog["utterances"][-1]["annotations"]["intent_catcher"][
intent_name] = not_detected
if "/new_persona" in user_uttr_text:
# process /new_persona command
skills_for_uttr.append(
"personality_catcher"
) # TODO: rm crutch of personality_catcher
elif user_uttr_text == "/get_dialog_id":
skills_for_uttr.append("dummy_skill")
elif high_priority_intent_detected:
# process intent with corresponding IntentResponder
skills_for_uttr.append("dff_intent_responder_skill")
elif is_sensitive_topic_and_request(user_uttr):
# process user utterance with sensitive content, "safe mode"
# adding open-domain skills without opinion expression
skills_for_uttr.append("dff_program_y_dangerous_skill")
skills_for_uttr.append("meta_script_skill")
skills_for_uttr.append("personal_info_skill")
skills_for_uttr.append("factoid_qa")
skills_for_uttr.append("dff_grounding_skill")
skills_for_uttr.append("dummy_skill")
skills_for_uttr.append("small_talk_skill")
if if_lets_chat_about_particular_topic_detected:
skills_for_uttr.append("news_api_skill")
if if_special_weather_turn_on(user_uttr, bot_uttr):
skills_for_uttr.append("dff_weather_skill")
if is_celebrity_mentioned:
skills_for_uttr.append("dff_gossip_skill")
# adding closed-domain skills
skills_for_uttr += turn_on_skills(
detected_topics,
intent_catcher_intents,
user_uttr_text,
bot_uttr.get("text", ""),
available_skills=[
"news_api_skill",
"dff_coronavirus_skill",
"dff_funfact_skill",
"dff_weather_skill",
"dff_short_story_skill",
],
)
# adding linked-to skills
skills_for_uttr.extend(get_linked_to_skills(dialog))
skills_for_uttr.extend(get_previously_active_skill(dialog))
else:
# general case
if low_priority_intent_detected:
skills_for_uttr.append("dff_intent_responder_skill")
# adding open-domain skills
skills_for_uttr.append("dff_grounding_skill")
skills_for_uttr.append("dff_program_y_skill")
skills_for_uttr.append("personal_info_skill")
skills_for_uttr.append("meta_script_skill")
skills_for_uttr.append("dummy_skill")
skills_for_uttr.append("dialogpt") # generative skill
skills_for_uttr.append("small_talk_skill")
skills_for_uttr.append("knowledge_grounding_skill")
skills_for_uttr.append("convert_reddit")
skills_for_uttr.append("comet_dialog_skill")
skills_for_uttr.append("dff_program_y_wide_skill")
# adding friendship only in the beginning of the dialog
if len(dialog["utterances"]) < 20:
skills_for_uttr.append("dff_friendship_skill")
if if_choose_topic_detected or if_lets_chat_about_particular_topic_detected:
skills_for_uttr.append("knowledge_grounding_skill")
skills_for_uttr.append("news_api_skill")
switch_wiki_skill, _ = if_switch_wiki_skill(
user_uttr, bot_uttr)
if switch_wiki_skill or switch_wiki_skill_on_news(
user_uttr, bot_uttr):
skills_for_uttr.append("dff_wiki_skill")
if if_switch_test_skill(user_uttr, bot_uttr):
skills_for_uttr.append("dff_art_skill")
# adding factoidQA Skill if user utterance is factoid question
if is_factoid:
skills_for_uttr.append("factoid_qa")
if "dummy_skill" in prev_active_skill and len(
dialog["utterances"]) > 4:
skills_for_uttr.append("dummy_skill_dialog")
# if user mentions
if is_celebrity_mentioned:
skills_for_uttr.append("dff_gossip_skill")
# some special cases
if if_special_weather_turn_on(user_uttr, bot_uttr):
skills_for_uttr.append("dff_weather_skill")
if if_turn_on_emotion(user_uttr, bot_uttr):
skills_for_uttr.append("emotion_skill")
if get_named_locations(user_uttr):
skills_for_uttr.append("dff_travel_skill")
if extract_movies_names_from_annotations(user_uttr):
skills_for_uttr.append("dff_movie_skill")
# adding closed-domain skills
skills_for_uttr += turn_on_skills(
detected_topics,
intent_catcher_intents,
user_uttr_text,
bot_uttr.get("text", ""),
available_skills=[
"dff_art_skill",
"dff_movie_skill",
"dff_book_skill",
"news_api_skill",
"dff_food_skill",
"dff_animals_skill",
"dff_sport_skill",
"dff_music_skill",
"dff_science_skill",
"dff_gossip_skill",
"game_cooperative_skill",
"dff_weather_skill",
"dff_funfact_skill",
"dff_travel_skill",
"dff_coronavirus_skill",
"dff_bot_persona_skill",
"dff_gaming_skill",
"dff_short_story_skill",
],
)
# adding linked-to skills
skills_for_uttr.extend(get_linked_to_skills(dialog))
skills_for_uttr.extend(get_previously_active_skill(dialog))
# NOW IT IS NOT ONLY FOR USUAL CONVERSATION BUT ALSO FOR SENSITIVE/HIGH PRIORITY INTENTS/ETC
if "dff_coronavirus_skill" in skills_for_uttr:
# no convert & comet when about coronavirus
if "convert_reddit" in skills_for_uttr:
skills_for_uttr.remove("convert_reddit")
if "comet_dialog_skill" in skills_for_uttr:
skills_for_uttr.remove("comet_dialog_skill")
if len(dialog["utterances"]) > 1:
# Use only misheard asr skill if asr is not confident and skip it for greeting
if user_uttr_annotations.get("asr",
{}).get("asr_confidence",
"high") == "very_low":
skills_for_uttr = ["misheard_asr"]
if "/alexa_" in user_uttr_text:
# adding alexa handler for Amazon Alexa specific commands
skills_for_uttr = ["alexa_handler"]
logger.info(f"Selected skills: {skills_for_uttr}")
total_time = time.time() - st_time
logger.info(f"rule_based_selector exec time = {total_time:.3f}s")
asyncio.create_task(
callback(task_id=payload["task_id"],
response=list(set(skills_for_uttr))))
except Exception as e:
total_time = time.time() - st_time
logger.info(f"rule_based_selector exec time = {total_time:.3f}s")
logger.exception(e)
sentry_sdk.capture_exception(e)
asyncio.create_task(
callback(task_id=payload["task_id"],
response=["dff_program_y_skill", "dummy_skill"]))
def tag_based_response_selection(dialog,
candidates,
scores,
confidences,
bot_utterances,
all_prev_active_skills=None):
all_prev_active_skills = all_prev_active_skills if all_prev_active_skills is not None else []
all_prev_active_skills = Counter(all_prev_active_skills)
annotated_uttr = dialog["human_utterances"][-1]
all_user_intents, all_user_topics, all_user_named_entities, all_user_nounphrases = get_main_info_annotations(
annotated_uttr)
_is_switch_topic_request = is_switch_topic(annotated_uttr)
_is_force_intent = any(
[_intent in all_user_intents for _intent in FORCE_INTENTS_IC.keys()])
# if user utterance contains any question (REGEXP & punctuation check!)
_is_require_action_intent = is_any_question_sentence_in_utterance({
"text":
annotated_uttr.get("annotations",
{}).get("sentseg", {}).get("punct_sent",
annotated_uttr["text"])
})
# if user utterance contains any question AND requires some intent by socialbot
_is_require_action_intent = _is_require_action_intent and any([
_intent in all_user_intents
for _intent in REQUIRE_ACTION_INTENTS.keys()
])
_force_intents_detected = [
_intent for _intent in FORCE_INTENTS_IC.keys()
if _intent in all_user_intents
]
# list of user intents which require some action by socialbot
_require_action_intents_detected = [
_intent for _intent in REQUIRE_ACTION_INTENTS.keys()
if _intent in all_user_intents
]
_force_intents_skills = sum([
FORCE_INTENTS_IC.get(_intent, [])
for _intent in _force_intents_detected
], [])
# list of intents required by the socialbot
_required_actions = sum([
REQUIRE_ACTION_INTENTS.get(_intent, [])
for _intent in _require_action_intents_detected
], [])
_contains_entities = len(
get_entities(annotated_uttr, only_named=False, with_labels=False)) > 0
_is_active_skill_can_not_continue = False
_prev_bot_uttr = dialog["bot_utterances"][-1] if len(
dialog["bot_utterances"]) > 0 else {}
_prev_active_skill = dialog["bot_utterances"][-1]["active_skill"] if len(
dialog["bot_utterances"]) > 0 else ""
_prev_prev_active_skill = dialog["bot_utterances"][-2][
"active_skill"] if len(dialog["bot_utterances"]) > 1 else ""
_no_script_two_times_in_a_row = False
if _prev_active_skill and _prev_prev_active_skill:
if all([
skill not in ACTIVE_SKILLS + ALMOST_ACTIVE_SKILLS
for skill in [_prev_active_skill, _prev_prev_active_skill]
]):
_no_script_two_times_in_a_row = True
disliked_skills = get_updated_disliked_skills(
dialog, can_not_be_disliked_skills=CAN_NOT_BE_DISLIKED_SKILLS)
_is_dummy_linkto_available = any([
cand_uttr["skill_name"] == "dummy_skill"
and cand_uttr.get("type", "") == "link_to_for_response_selector"
for cand_uttr in candidates
])
categorized_hyps = {}
categorized_prompts = {}
for dasuffix in ["reqda", ""]:
for actsuffix in ["active", "continued", "finished"]:
for suffix in [
"same_topic_entity_no_db",
"same_topic_entity_db",
"othr_topic_entity_no_db",
"othr_topic_entity_db",
]:
categorized_hyps[f"{actsuffix}_{suffix}_{dasuffix}"] = []
categorized_prompts[f"{actsuffix}_{suffix}_{dasuffix}"] = []
CASE = ""
acknowledgement_hypothesis = {}
for cand_id, cand_uttr in enumerate(candidates):
if confidences[cand_id] == 0.0 and cand_uttr[
"skill_name"] not in ACTIVE_SKILLS:
logger.info(
f"Dropping cand_id: {cand_id} due to toxicity/badlists")
continue
all_cand_intents, all_cand_topics, all_cand_named_entities, all_cand_nounphrases = get_main_info_annotations(
cand_uttr)
skill_name = cand_uttr["skill_name"]
_is_dialog_abandon = get_dialog_breakdown_annotations(
cand_uttr) and PRIORITIZE_NO_DIALOG_BREAKDOWN
_is_just_prompt = (cand_uttr["skill_name"] == "dummy_skill" and any([
question_type in cand_uttr.get("type", "") for question_type in
["normal_question", "link_to_for_response_selector"]
])) or cand_uttr.get("response_parts", []) == ["prompt"]
if cand_uttr["confidence"] == 1.0:
# for those hypotheses where developer forgot to set tag to MUST_CONTINUE
cand_uttr["can_continue"] = MUST_CONTINUE
_can_continue = cand_uttr.get("can_continue", CAN_NOT_CONTINUE)
_user_wants_to_chat_about_topic = (
if_chat_about_particular_topic(annotated_uttr)
and "about it" not in annotated_uttr["text"].lower())
_user_does_not_want_to_chat_about_topic = if_not_want_to_chat_about_particular_topic(
annotated_uttr)
_user_wants_bot_to_choose_topic = if_choose_topic(
annotated_uttr, _prev_bot_uttr)
if any([
phrase.lower() in cand_uttr["text"].lower()
for phrase in LINK_TO_PHRASES
]):
# add `prompt` to response_parts if any linkto phrase in hypothesis
cand_uttr["response_parts"] = cand_uttr.get("response_parts",
[]) + ["prompt"]
# identifies if candidate contains named entities from last human utterance
_same_named_entities = (len(
get_common_tokens_in_lists_of_strings(
all_cand_named_entities, all_user_named_entities)) > 0)
# identifies if candidate contains all (not only named) entities from last human utterance
_same_nounphrases = len(
get_common_tokens_in_lists_of_strings(all_cand_nounphrases,
all_user_nounphrases)) > 0
_same_topic_entity = (_same_named_entities or _same_nounphrases
) and PRIORITIZE_WITH_SAME_TOPIC_ENTITY
_is_active_skill = (_prev_active_skill == cand_uttr["skill_name"] or
cand_uttr.get("can_continue", "") == MUST_CONTINUE)
_is_active_skill = _is_active_skill and skill_name in ACTIVE_SKILLS
_is_active_skill = _is_active_skill and (_can_continue in [
MUST_CONTINUE, CAN_CONTINUE_SCENARIO, CAN_NOT_CONTINUE
] or (_can_continue == CAN_CONTINUE_PROMPT
and all_prev_active_skills.get(skill_name, []) < 10))
_is_active_skill = _is_active_skill and PRIORITIZE_SCRIPTED_SKILLS
if _is_active_skill:
# we will forcibly add prompt if current scripted skill finishes scenario,
# and has no opportunity to continue at all.
_is_active_skill_can_not_continue = _is_active_skill and _can_continue in [
CAN_NOT_CONTINUE
]
if _is_force_intent:
# =====force intents, choose as best_on_topic hypotheses from skills responding this request=====
CASE = "Force intent."
if cand_uttr["skill_name"] in _force_intents_skills:
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=False,
)
elif _is_switch_topic_request or _user_does_not_want_to_chat_about_topic or _user_wants_bot_to_choose_topic:
# =====direct request by user to switch the topic of current conversation=====
# give priority to dummy linkto hypothesis if available, else other prompts if available.
_is_active_skill = (
cand_uttr.get("type", "") == "link_to_for_response_selector"
if _is_dummy_linkto_available else _is_just_prompt)
# no priority to must_continue to skip incorrect continuation of script
_can_continue = CAN_CONTINUE_SCENARIO if _can_continue == MUST_CONTINUE else _can_continue
CASE = "Switch topic intent."
if len(all_user_named_entities) > 0 or len(
all_user_nounphrases) > 0:
# -----user defines new topic/entity-----
# _same_topic_entity does not depend on hyperparameter in these case
_same_topic_entity = _same_named_entities or _same_nounphrases
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=False,
)
else:
# -----user want socialbot to define new topic/entity-----
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=False,
)
elif _user_wants_to_chat_about_topic:
# user wants to chat about particular topic
CASE = "User wants to talk about topic."
# in this case we do not give priority to previously active skill (but give to must continue skill!)
# because now user wants to talk about something particular
_is_active_skill = cand_uttr.get("can_continue",
"") == MUST_CONTINUE
# _same_topic_entity does not depend on hyperparameter in these case
_same_topic_entity = _same_named_entities or _same_nounphrases
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=False,
)
elif _is_require_action_intent and PRIORITIZE_WITH_REQUIRED_ACT:
# =====user intent requires particular action=====
CASE = "User intent requires action. USER UTTERANCE CONTAINS QUESTION."
_is_grounding_reqda = (skill_name == "dff_grounding_skill"
and cand_uttr.get(
"type", "") == "universal_response")
_is_active_skill = cand_uttr.get(
"can_continue",
"") == MUST_CONTINUE # no priority to prev active skill
_can_continue = CAN_NOT_CONTINUE # no priority to scripted skills
if set(all_cand_intents).intersection(
set(_required_actions
)) or _is_grounding_reqda or _is_active_skill:
# -----one of the can intent is in intents required by user-----
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=True,
)
else:
# -----NO required dialog acts-----
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=False,
)
else:
# =====user intent does NOT require particular action=====
CASE = "General case."
categorized_hyps, categorized_prompts = categorize_candidate(
cand_id,
skill_name,
categorized_hyps,
categorized_prompts,
_is_just_prompt,
_is_active_skill,
_can_continue,
_same_topic_entity,
_is_dialog_abandon,
_is_required_da=False,
)
# a bit of rule based help
if (len(dialog["human_utterances"]) == 1
and cand_uttr["skill_name"] == "dff_friendship_skill"
and greeting_spec in cand_uttr["text"]):
categorized_hyps = add_to_top1_category(cand_id, categorized_hyps,
_is_require_action_intent)
elif (cand_uttr["skill_name"] == "dff_friendship_skill"
and (how_are_you_spec in cand_uttr["text"]
or what_i_can_do_spec in cand_uttr["text"])
and len(dialog["utterances"]) < 16):
categorized_hyps = add_to_top1_category(cand_id, categorized_hyps,
_is_require_action_intent)
# elif cand_uttr["skill_name"] == 'program_y_dangerous' and cand_uttr['confidence'] == 0.98:
# categorized_hyps = add_to_top1_category(cand_id, categorized_hyps, _is_require_action_intent)
elif cand_uttr[
"skill_name"] == "small_talk_skill" and is_sensitive_situation(
dialog["human_utterances"][-1]):
# let small talk to talk about sex ^_^
categorized_hyps = add_to_top1_category(cand_id, categorized_hyps,
_is_require_action_intent)
elif cand_uttr["confidence"] >= 1.0:
# -------------------- SUPER CONFIDENCE CASE HERE! --------------------
categorized_hyps = add_to_top1_category(cand_id, categorized_hyps,
_is_require_action_intent)
if cand_uttr["skill_name"] == "dff_grounding_skill" and [
"acknowledgement"
] == cand_uttr.get("response_parts", []):
acknowledgement_hypothesis = deepcopy(cand_uttr)
logger.info(f"Current CASE: {CASE}")
# now compute current scores as one float value
curr_single_scores = compute_curr_single_scores(candidates, scores,
confidences)
# remove disliked skills from hypotheses
if IGNORE_DISLIKED_SKILLS:
for category in categorized_hyps:
new_ids = []
for cand_id in categorized_hyps[category]:
if (candidates[cand_id]["skill_name"] in disliked_skills
and candidates[cand_id].get("can_continue",
CAN_NOT_CONTINUE)
== MUST_CONTINUE):
disliked_skills.remove(candidates[cand_id]["skill_name"])
if candidates[cand_id]["skill_name"] not in disliked_skills:
new_ids.append(cand_id)
categorized_hyps[category] = deepcopy(new_ids)
for category in categorized_prompts:
new_ids = []
for cand_id in categorized_prompts[category]:
if (candidates[cand_id]["skill_name"] in disliked_skills
and candidates[cand_id].get("can_continue",
CAN_NOT_CONTINUE)
== MUST_CONTINUE):
disliked_skills.remove(candidates[cand_id]["skill_name"])
if candidates[cand_id]["skill_name"] not in disliked_skills:
new_ids.append(cand_id)
categorized_prompts[category] = deepcopy(new_ids)
best_cand_id = pickup_best_id(categorized_hyps, candidates,
curr_single_scores, bot_utterances)
best_candidate = candidates[best_cand_id]
best_candidate["human_attributes"] = best_candidate.get(
"human_attributes", {})
# save updated disliked skills to human attributes of the best candidate
best_candidate["human_attributes"]["disliked_skills"] = disliked_skills
logger.info(f"Best candidate: {best_candidate}")
n_sents_without_prompt = len(sent_tokenize(best_candidate["text"]))
_is_best_not_script = best_candidate[
"skill_name"] not in ACTIVE_SKILLS + ALMOST_ACTIVE_SKILLS
no_question_by_user = "******" not in dialog["human_utterances"][-1][
"annotations"].get("sentseg",
{}).get("punct_sent",
dialog["human_utterances"][-1]["text"])
# if `no` to 1st in a row linkto question, and chosen response is not from scripted skill
_no_to_first_linkto = is_no(dialog["human_utterances"][-1]) and any([
phrase.lower() in _prev_bot_uttr.get("text", "").lower()
for phrase in LINK_TO_PHRASES
])
# if chosen short response or question by not-scripted skill
_is_short_or_question_by_not_script = _is_best_not_script and (
"?" in best_candidate["text"]
or len(best_candidate["text"].split()) < 4)
_no_questions_for_3_steps = not any([
is_any_question_sentence_in_utterance(uttr)
for uttr in dialog["bot_utterances"][-3:]
])
if PRIORITIZE_PROMTS_WHEN_NO_SCRIPTS:
if (_no_script_two_times_in_a_row
and _is_short_or_question_by_not_script
and no_question_by_user) or (_no_to_first_linkto
and _is_best_not_script):
# if no scripted skills 2 time sin a row before, current chosen best cand is not scripted, contains `?`,
# and user utterance does not contain "?", replace utterance with dummy!
best_prompt_id = pickup_best_id(categorized_prompts, candidates,
curr_single_scores, bot_utterances)
best_candidate = deepcopy(candidates[best_prompt_id])
best_cand_id = best_prompt_id
if does_not_require_prompt(candidates, best_cand_id):
# the candidate already contains a prompt or a question or of a length more than 200 symbols
logger.info(
"Best candidate contains prompt, question, request or length of > 200 symbols. Do NOT add prompt."
)
pass
elif sum(categorized_prompts.values(), []):
# best cand is 3d times in a row not scripted skill, let's append linkto
# need to add some prompt, and have a prompt
_add_prompt_forcibly = best_candidate[
"skill_name"] == _prev_active_skill and _is_active_skill_can_not_continue
_add_prompt_forcibly = _add_prompt_forcibly and not _contains_entities
# prompts are added:
# - in 1 out of 10 cases, if current human utterance does not contain entities,
# and no prompt for several last bot utterances
# - if PRIORITIZE_PROMTS_WHEN_NO_SCRIPTS and current utterance is from active on prev step scripted skill and
# it has a status can-not-continue
# - if PRIORITIZE_PROMTS_WHEN_NO_SCRIPTS and last 2 bot uttr are not from scripted skill,
# and current best uttr is also from not-scripted skill
if ((prompt_decision() and not _contains_entities
and _no_questions_for_3_steps)
or (_add_prompt_forcibly and PRIORITIZE_PROMTS_WHEN_NO_SCRIPTS)
or
(PRIORITIZE_PROMTS_WHEN_NO_SCRIPTS
and _no_script_two_times_in_a_row and _is_best_not_script)):
logger.info("Decided to add a prompt to the best candidate.")
best_prompt_id = pickup_best_id(categorized_prompts, candidates,
curr_single_scores, bot_utterances)
# as we have only one active skill, let's consider active skill as that one providing prompt
# but we also need to reassign all the attributes
best_prompt = candidates[best_prompt_id]
best_candidate[
"text"] = f'{best_candidate["text"]} {best_prompt["text"]}'
best_candidate["attributes"] = best_candidate.get("attributes", {})
best_candidate["attributes"]["prompt_skill"] = best_prompt
# anyway we must combine used links
best_candidate["human_attributes"] = best_candidate.get(
"human_attributes", {})
best_candidate["human_attributes"] = join_used_links_in_attributes(
best_candidate["human_attributes"],
best_prompt.get("human_attributes", {}))
if len(best_candidate["human_attributes"]["used_links"]) == 0:
best_candidate["human_attributes"].pop("used_links")
was_ackn = if_acknowledgement_in_previous_bot_utterance(dialog)
best_resp_cont_ackn = "acknowledgement" in best_candidate.get(
"response_parts", [])
if (ADD_ACKNOWLEDGMENTS_IF_POSSIBLE and acknowledgement_hypothesis
and acknowledgement_decision(all_user_intents)
and n_sents_without_prompt == 1 and not was_ackn
and not best_resp_cont_ackn):
logger.info(
"Acknowledgement is given, Final hypothesis contains only 1 sentence, no ackn in prev bot uttr,"
"and we decided to add an acknowledgement to the best candidate.")
best_candidate[
"text"] = f'{acknowledgement_hypothesis["text"]} {best_candidate["text"]}'
best_candidate["response_parts"] = ["acknowledgement"
] + best_candidate.get(
"response_parts", [])
return best_candidate, best_cand_id, curr_single_scores
def get_response_for_particular_topic_and_status(topic,
curr_meta_script_status,
dialog, source_topic):
attr = {
"meta_script_topic": topic,
"meta_script_status": curr_meta_script_status
}
if len(dialog["human_utterances"]) > 0:
user_uttr = dialog["human_utterances"][-1]
text_user_uttr = dialog["human_utterances"][-1]["text"].lower()
last_user_sent_text = (dialog["human_utterances"][-1].get(
"annotations", {}).get("sentseg", {}).get("segments",
[""])[-1].lower())
else:
user_uttr = {"text": ""}
text_user_uttr = ""
last_user_sent_text = ""
if len(dialog["bot_utterances"]) > 0:
bot_uttr = dialog["bot_utterances"][-1]
else:
bot_uttr = {}
if curr_meta_script_status == "starting":
response, confidence, attr = get_starting_phrase(dialog, topic, attr)
attr["response_parts"] = ["prompt"]
can_offer_topic = if_choose_topic(dialog["human_utterances"][-1],
bot_uttr)
talk_about_user_topic = is_custom_topic(
topic) and if_chat_about_particular_topic(user_uttr, bot_uttr)
prev_what_to_talk_about_outputs = [
get_outputs_with_response_from_dialog(dialog["utterances"][-3:],
response=response,
activated=True) for response
in GREETING_QUESTIONS[list(GREETING_QUESTIONS.keys())[0]]
]
prev_what_to_talk_about_outputs = sum([
list_of_outputs
for list_of_outputs in prev_what_to_talk_about_outputs
if len(list_of_outputs) > 0
], [])
prev_what_to_talk_about_greeting = len(
prev_what_to_talk_about_outputs) > 0 and bot_uttr.get(
"active_skill", "") in ["dff_friendship_skill", "program_y"]
if (not prev_what_to_talk_about_greeting
and can_offer_topic) or talk_about_user_topic:
# if person wants to talk about something particular and we have extracted some topic - do that!
confidence = MATCHED_DIALOG_BEGIN_CONFIDENCE
elif "?" in last_user_sent_text or prev_what_to_talk_about_greeting:
# if some question was asked by user, do not start script at all!
response, confidence = "", 0.0
elif len(dialog["utterances"]) <= 20:
confidence = DEFAULT_DIALOG_BEGIN_CONFIDENCE
elif source_topic == NP_SOURCE:
confidence = NOUN_TOPIC_STARTING_CONFIDENCE
else:
confidence = DEFAULT_STARTING_CONFIDENCE
else:
if curr_meta_script_status == "deeper1" and "?" in last_user_sent_text and "what" not in text_user_uttr:
response, confidence, attr = "", 0.0, {}
elif "?" in last_user_sent_text and not check_topic_lemmas_in_sentence(
text_user_uttr, topic):
logger.info(
"Question by user was detected. Without any word from topic in it. "
"Don't continue the script on this turn.")
response, confidence, attr = "", 0.0, {}
elif is_switch_topic(user_uttr) or if_chat_about_particular_topic(
user_uttr):
logger.info("Topic switching was detected. Finish script.")
response, confidence = FINISHED_SCRIPT_RESPONSE, 0.5
attr["meta_script_status"] = FINISHED_SCRIPT
attr["can_continue"] = CAN_NOT_CONTINUE
elif get_user_replies_to_particular_skill(
dialog["utterances"],
"meta_script_skill")[-2:] == ["no.", "no."]:
logger.info(
"Two consequent `no` answers were detected. Finish script.")
response, confidence = FINISHED_SCRIPT_RESPONSE, 0.5
attr["meta_script_status"] = FINISHED_SCRIPT
attr["can_continue"] = CAN_NOT_CONTINUE
elif curr_meta_script_status == "comment":
response, confidence, attr = get_comment_phrase(dialog, attr)
attr["can_continue"] = CAN_NOT_CONTINUE
elif curr_meta_script_status == "opinion":
response, confidence, attr = get_opinion_phrase(
dialog, topic, attr)
elif curr_meta_script_status == "deeper1" and (
is_no(user_uttr) or "never" in text_user_uttr):
response, confidence = FINISHED_SCRIPT_RESPONSE, 0.5
attr["meta_script_status"] = FINISHED_SCRIPT
attr["can_continue"] = CAN_NOT_CONTINUE
else:
response, confidence, attr = get_statement_phrase(
dialog, topic, attr, TOPICS)
attr["can_continue"] = CAN_CONTINUE_SCENARIO
if confidence > 0.7 and (is_yes(user_uttr)
or len(text_user_uttr.split()) > 7):
# if yes detected, confidence 1.0 - we like agreements!
confidence = 1.0
if confidence > 0.7 and bot_uttr.get("active_skill",
"") != "meta_script_skill":
confidence = BROKEN_DIALOG_CONTINUE_CONFIDENCE
logger.info(
f"User sent: `{text_user_uttr}`. Response: `{response}`. Attr: `{attr}.`"
)
return response, confidence, attr
def get_statuses_and_topics(dialog):
"""
Find prevously discussed meta-script topics, the last met-script status,
determine current step meta-script status and topic.
Args:
dialog: dialog itself
Returns:
tuple of current status and topic
"""
# deeper2 and opinion could be randomly skipped in dialog flow
dialog_flow = ["starting", "deeper1", "deeper2", "opinion", "comment"]
dialog_flow_user_topic = ["starting", "deeper1", "comment"]
curr_meta_script_statuses = []
curr_meta_script_topics = []
source_topics = []
if len(dialog["utterances"]) >= 3:
# if dialog is not empty
used_topics = get_used_attributes_by_name(
dialog["utterances"],
attribute_name="meta_script_topic",
value_by_default="",
activated=True)
# this determines how many replies back we assume active meta script skill to continue dialog.
# let's assume we can continue if meta_scrip skill was active on up to 2 steps back
prev_reply_output = get_skill_outputs_from_dialog(
dialog["utterances"][-5:],
skill_name="meta_script_skill",
activated=True)
# get last meta script output even if it was not activated but right after it was active
last_all_meta_script_outputs = get_skill_outputs_from_dialog(
dialog["utterances"][-5:],
skill_name="meta_script_skill",
activated=False)
prev_topic_finished = False
for out in last_all_meta_script_outputs:
if out.get("meta_script_status", "") == "finished":
logger.info(
f"Found finished dialog on meta_script_topic: `{out.get('meta_script_status', '')}`"
)
prev_topic_finished = True
if len(prev_reply_output) > 0:
# previously active skill was `meta_script_skill`
curr_meta_script_status = prev_reply_output[-1].get(
"meta_script_status", "")
else:
# previous active skill was not `meta_script_skill`
curr_meta_script_status = ""
logger.info(f"Found meta_script_status: `{curr_meta_script_status}`")
if curr_meta_script_status in ["comment", "", FINISHED_SCRIPT
] or prev_topic_finished:
# if previous meta script is finished (comment given) in previous bot reply
# or if no meta script in previous reply or script was forcibly
topics, curr_source_topics = get_not_used_topics(
used_topics, dialog)
if curr_source_topics != [PREDEFINED_SOURCE]:
# if topic is extracted from utterances
pass
elif if_choose_topic(dialog["human_utterances"][-1],
dialog["bot_utterances"][-1]):
# len(utterances) >3 so at least 1 bot utterance exists
# one of the predefined topics (wiki or hand-written)
curr_meta_script_statuses += [dialog_flow[0]] * len(topics)
curr_meta_script_topics += topics
source_topics += curr_source_topics
else:
pass
else:
# some meta script is already in progress
# we define it here as predefined because we do not care about this variable if it's not script starting
source_topic = PREDEFINED_SOURCE
curr_meta_script_topic = used_topics[-1]
logger.info(
f"Found meta_script_status: `{curr_meta_script_status}` "
f"on previous meta_script_topic: `{curr_meta_script_topic}`")
# getting the next dialog flow status
if is_custom_topic(curr_meta_script_topic):
curr_meta_script_status = dialog_flow_user_topic[
dialog_flow_user_topic.index(curr_meta_script_status) + 1]
else:
curr_meta_script_status = dialog_flow[
dialog_flow.index(curr_meta_script_status) + 1]
if curr_meta_script_status == "opinion":
# randomly skip third deeper question
if uniform(0, 1) <= 0.5:
curr_meta_script_status = "comment"
if curr_meta_script_status == "deeper2":
# randomly skip third deeper question
if uniform(0, 1) <= 0.5:
curr_meta_script_status = "opinion"
logger.info(f"New meta_script_status: `{curr_meta_script_status}` "
f"on meta_script_topic: `{curr_meta_script_topic}`")
curr_meta_script_statuses += [curr_meta_script_status]
curr_meta_script_topics += [curr_meta_script_topic]
source_topics += [source_topic]
else:
# start of the dialog, pick up a topic of meta script
curr_meta_script_topics, source_topics = get_not_used_topics([],
dialog)
if source_topics != [PREDEFINED_SOURCE]:
curr_meta_script_statuses = [dialog_flow_user_topic[0]
] * len(curr_meta_script_topics)
else:
curr_meta_script_statuses = [dialog_flow[0]
] * len(curr_meta_script_topics)
logger.info(f"Final new meta_script_status: `{curr_meta_script_statuses}` "
f"on meta_script_topic: `{curr_meta_script_topics}`")
return curr_meta_script_statuses, curr_meta_script_topics, source_topics
def rule_score_based_selection(dialog, candidates, scores, confidences, is_toxics, bot_utterances):
curr_single_scores = []
bot_utt_counter = Counter(bot_utterances)
lower_duplicates_score(candidates, bot_utt_counter, scores, confidences)
lower_retrieve_skills_confidence_if_scenario_exist(candidates, scores, confidences)
# prev_active_skill = dialog["bot_utterances"][-1]['active_skill'] if len(dialog["bot_utterances"]) > 0 else ''
skill_names = [c["skill_name"] for c in candidates]
very_big_score = 100
very_low_score = -100
dummy_question = ""
dummy_question_human_attr = {}
link_to_question = ""
link_to_human_attrs = {}
not_sure_factoid = False
if "factoid_qa" in skill_names:
factoid_index = skill_names.index("factoid_qa")
logging.debug("factoid")
logging.debug(str(candidates[factoid_index]))
if "not sure" in candidates[factoid_index] and candidates[factoid_index]["not sure"]:
not_sure_factoid = True
for i in range(len(scores)):
curr_score = None
is_misheard = misheard_with_spec1 in candidates[i]["text"] or misheard_with_spec2 in candidates[i]["text"]
intent_name = get_intent_name(candidates[i]["text"])
is_intent_candidate = (skill_names[i] in ["dff_intent_responder_skill", "dff_program_y_skill"]) and intent_name
is_intent_candidate = is_intent_candidate and intent_name not in low_priority_intents
# print("is intent candidate? " + str(is_intent_candidate), flush=True)
if len(dialog["human_utterances"]) == 1 and greeting_spec not in candidates[i]["text"]:
logger.info("Dialog Beginning detected.")
if (
if_chat_about_particular_topic(dialog["utterances"][0])
and "about it" not in dialog["utterances"][0]["text"].lower()
):
logger.info("User wants to talk about particular topic")
# if user says `let's chat about blablabla`
if skill_names[i] == "factoid_qa":
logger.info("Particular topic. Facts + Greeting to very big score.")
# I don't have an opinion on that but I know some facts.
resp = candidates[i]["text"].replace("I don't have an opinion on that but I know some facts.", "")
candidates[i]["text"] = "Hi, " + greeting_spec + "! " + resp
curr_score = very_big_score
elif skill_names[i] == "meta_script_skill" and len(candidates[i]["text"]) > 0 and confidences[i] > 0.98:
logger.info("Particular topic. meta_script_skill + Greeting to very big score.")
# I don't have an opinion on that but I know some facts.
resp = candidates[i]["text"]
candidates[i]["text"] = "Hi, " + greeting_spec + "! " + resp
curr_score = very_big_score
elif skill_names[i] == "small_talk_skill":
logger.info("Particular topic. Small-talk + Greeting NOT to very big score.")
# for now do not give small talk a very big score here
candidates[i]["text"] = "Hi, " + greeting_spec + "! " + candidates[i]["text"]
# curr_score = very_big_score
elif if_choose_topic(dialog["utterances"][0]) and "about it" not in dialog["utterances"][0]["text"].lower():
logger.info("User wants bot to choose the topic")
# if user says `let's chat about something`
if skill_names[i] == "small_talk_skill":
logger.info("No topic. Small-talk + Greeting to very big score.")
candidates[i]["text"] = "Hi, " + greeting_spec + "! " + candidates[i]["text"]
curr_score = very_big_score
elif skill_names[i] == "meta_script_skill" and len(candidates[i]["text"]) > 0:
logger.info("No topic. Meta-script + Greeting to very big score.")
candidates[i]["text"] = "Hi, " + greeting_spec + "! " + candidates[i]["text"]
curr_score = very_big_score
else:
logger.info("User just wants to talk.")
# if user says something else
if skill_names[i] == "program_y" and greeting_spec in candidates[i]["text"]:
logger.info("Just chat. Program-y to very big score.")
curr_score = very_big_score
elif (
skill_names[i] == "dff_friendship_skill"
and (how_are_you_spec in candidates[i]["text"] or what_i_can_do_spec in candidates[i]["text"])
and len(dialog["utterances"]) < 16
):
curr_score = very_big_score
elif skill_names[i] == "dff_friendship_skill" and greeting_spec in candidates[i]["text"]:
if len(dialog["utterances"]) < 2:
curr_score = very_big_score
else:
confidences[i] = 0.2 # Low confidence for greeting in the middle of dialogue
# we don't have 'cobotqa' anymore; instead we have factoid_qa
elif skill_names[i] in ["factoid_qa"] and "Here's something I found on the web." in candidates[i]["text"]:
confidences[i] = 0.6
elif (
skill_names[i] == "factoid_qa"
and dialog["human_utterances"][-1]["annotations"]
.get("intent_catcher", {})
.get("weather_forecast_intent", {})
.get("detected", 0)
== 1
):
confidences[i] = 0.8
elif skill_names[i] == "misheard_asr" and is_misheard:
curr_score = very_big_score
elif is_intent_candidate:
curr_score = very_big_score
elif skill_names[i] in ["dummy_skill", "convert_reddit", "alice", "eliza", "tdidf_retrieval", "program_y"]:
if "question" in candidates[i].get("type", "") or "?" in candidates[i]["text"]:
penalty_start_utt = 1
if skill_names[i] == "program_y":
penalty_start_utt = 4
n_questions = 0
if len(bot_utterances) >= penalty_start_utt and "?" in bot_utterances[-1]:
confidences[i] /= 1.5
n_questions += 1
if len(bot_utterances) >= penalty_start_utt + 1 and "?" in bot_utterances[-2]:
confidences[i] /= 1.1
n_questions += 1
if n_questions == 2:
# two subsequent questions (1 / (1.5 * 1.1 * 1.2) = ~0.5)
confidences[i] /= 1.2
# this is only about `dummy_skill`
if "link_to_for_response_selector" in candidates[i].get("type", ""):
link_to_question = candidates[i]["text"]
link_to_human_attrs = candidates[i].get("human_attributes", {})
if skill_names[i] == "dummy_skill" and "question" in candidates[i].get("type", ""):
dummy_question = candidates[i]["text"]
dummy_question_human_attr = candidates[i].get("human_attributes", {})
if curr_score is None:
cand_scores = scores[i]
confidence = confidences[i]
skill_name = skill_names[i]
score_conv_eval = calculate_single_convers_evaluator_score(cand_scores)
score = CONV_EVAL_STRENGTH * score_conv_eval + CONFIDENCE_STRENGTH * confidence
logger.info(
f"Skill {skill_name} has final score: {score}. Confidence: {confidence}. "
f"Toxicity: {is_toxics[i]}. Cand scores: {cand_scores}"
)
curr_single_scores.append(score)
else:
cand_scores = scores[i]
skill_name = skill_names[i]
score_conv_eval = calculate_single_convers_evaluator_score(cand_scores)
score = CONV_EVAL_STRENGTH * score_conv_eval + curr_score
logger.info(
f"Skill {skill_name} has final score: {score}. " f"Toxicity: {is_toxics[i]}. Cand scores: {cand_scores}"
)
curr_single_scores.append(score)
highest_conf_exist = True if any(confidences >= 1.0) else False
if highest_conf_exist:
logger.info("Found skill with the highest confidence.")
for j in range(len(candidates)):
if highest_conf_exist and confidences[j] < 1.0 and curr_single_scores[j] < very_big_score:
# need to drop this candidates
logger.info(f"Dropping {skill_names[j]} which does not have a highest confidence or `very big score`.")
curr_single_scores[j] = very_low_score
best_id = np.argmax(curr_single_scores)
best_candidate = candidates[best_id]
best_skill_name = skill_names[int(best_id)]
best_candidate = add_question_to_statement(
best_candidate,
best_skill_name,
dummy_question,
dummy_question_human_attr,
link_to_question,
link_to_human_attrs,
not_sure_factoid,
)
return best_candidate, best_id, curr_single_scores
def respond():
print("response generation started")
st_time = time.time()
dialogs_batch = request.json["dialogs"]
# following 3 lists have len = number of samples going to the model
annotations_depths = []
dial_ids = []
input_batch = []
# following 4 lists have len = len(dialogs_batch)
entities = []
lets_chat_about_flags = []
nounphrases = []
special_intents_flags = []
chosen_topics = {}
for d_id, dialog in enumerate(dialogs_batch):
try:
user_input_text = dialog["human_utterances"][-1]["text"]
bot_uttr = dialog["bot_utterances"][-1] if len(
dialog["bot_utterances"]) > 0 else {}
switch_choose_topic = if_choose_topic(
dialog["human_utterances"][-1], bot_uttr)
# spacy_nounphrases
spacy_nounphrases = get_spacy_nounphrases(
dialog["human_utterances"][-1])
nounphrases.append(
re.compile(join_sentences_in_or_pattern(spacy_nounphrases), re.
IGNORECASE) if spacy_nounphrases else "")
# entities
curr_ents = get_named_entities(dialog["human_utterances"][-1])
entities.append(
re.compile(join_sentences_in_or_pattern(curr_ents), re.
IGNORECASE) if curr_ents else "")
# intents
lets_chat_about_flag, special_intents_flag = get_intents_flags(
dialog["human_utterances"][-1])
lets_chat_about_flags.append(lets_chat_about_flag)
special_intents_flags.append(special_intents_flag)
anntr_history_len = DEFAULT_ANNTR_HISTORY_LEN
bot_uttrs_for_dff_check = dialog["bot_utterances"][-2:] if len(
dialog["bot_utterances"]) > 1 else []
dffs_flag = check_dffs(bot_uttrs_for_dff_check)
if lets_chat_about_flag or switch_choose_topic:
anntr_history_len = 0
elif dffs_flag:
anntr_history_len = DFF_ANNTR_HISTORY_LEN
# if detected lets_chat is about topic from the file
lets_chat_topic = get_lets_chat_topic(
lets_chat_about_flag, dialog["human_utterances"][-1])
# if prev skill == news_api_skill get news description and create knowledge fact
news_api_fact = get_news_api_fact(
bot_uttr, dialog["human_utterances"],
not (switch_choose_topic or lets_chat_about_flag))
# start creating data for kg service
user_input_history = "\n".join(
[i["text"] for i in dialog["utterances"]])
annotators = {
# "odqa": "answer_sentence",
# "kbqa": "answer"
}
if not switch_choose_topic:
user_input_knowledge, annotations_depth = get_knowledge_from_annotators(
annotators, dialog["utterances"], anntr_history_len)
else:
user_input_knowledge = ""
annotations_depth = {}
# add nounphrases and entities to the knowledge
if user_input_knowledge:
user_input_checked_sentence = (
space_join(spacy_nounphrases) + space_join(curr_ents) +
tokenize.sent_tokenize(user_input_knowledge)[0])
else:
user_input_checked_sentence = ""
if user_input_knowledge:
user_input = {
"checked_sentence": user_input_checked_sentence,
"knowledge": user_input_knowledge,
"text": user_input_text,
"history": user_input_history,
}
annotations_depths.append(annotations_depth)
dial_ids.append(d_id)
input_batch.append(user_input)
retrieved_facts = get_annotations_from_dialog(
dialog["utterances"][-anntr_history_len * 2 - 1:],
"fact_retrieval")
if retrieved_facts:
for depth, fact in retrieved_facts[-TOP_N_FACTS:]:
user_input = {
"checked_sentence": fact,
"knowledge": fact,
"text": user_input_text,
"history": user_input_history,
}
input_batch.append(user_input)
annotations_depths.append({"retrieved_fact": depth})
dial_ids.append(d_id)
if any(
[switch_choose_topic, lets_chat_topic, lets_chat_about_flag]):
if lets_chat_topic:
fact = random.sample(TOPICS_FACTS[lets_chat_topic], 1)[0]
chosen_topics[d_id] = lets_chat_topic
_chosen_topic_fact = "lets_chat_cobot_da"
elif not get_entities(dialog["human_utterances"][-1],
only_named=False,
with_labels=False):
topic = random.sample(TOPICS_FACTS.keys(), 1)[0]
fact = random.sample(TOPICS_FACTS[topic], 1)[0]
chosen_topics[d_id] = topic
_chosen_topic_fact = "switch_random"
else:
fact = ""
if fact:
user_input = {
"checked_sentence": fact,
"knowledge": fact,
"text": user_input_text,
"history": user_input_history,
"chosen_topic_fact": _chosen_topic_fact,
}
input_batch.append(user_input)
annotations_depths.append({})
dial_ids.append(d_id)
if news_api_fact:
user_input = {
"checked_sentence": news_api_fact,
"knowledge": news_api_fact,
"text": user_input_text,
"history": user_input_history,
"news_api_fact": True,
}
input_batch.append(user_input)
annotations_depths.append({})
dial_ids.append(d_id)
fact_random_facts = get_fact_random(
dialog["utterances"][-anntr_history_len * 2 - 1:])
if fact_random_facts:
user_input = {
"checked_sentence": fact_random_facts[-1][1],
"knowledge": fact_random_facts[-1][1],
"text": user_input_text,
"history": user_input_history,
"fact_random_fact": True,
}
input_batch.append(user_input)
annotations_depths.append(
{"fact_random": fact_random_facts[-1][0]})
dial_ids.append(d_id)
user_news = get_news(dialog["human_utterances"][-1], "human")
bot_news = get_news(dialog["human_utterances"][-1], "bot")
# all_news = get_news(dialog["human_utterances"][-1], "all")
if user_news:
news_desc = user_news[-1].get("decsription", "")
if news_desc:
user_input = {
"checked_sentence": news_desc,
"knowledge": news_desc,
"text": user_input_text,
"history": user_input_history,
"news_fact": "human ",
}
input_batch.append(user_input)
annotations_depths.append({})
dial_ids.append(d_id)
elif bot_news:
news_desc = bot_news[-1].get("decsription", "")
if news_desc:
user_input = {
"checked_sentence": news_desc,
"knowledge": news_desc,
"text": user_input_text,
"history": user_input_history,
"news_fact": "bot ",
}
input_batch.append(user_input)
annotations_depths.append({})
dial_ids.append(d_id)
# elif all_news:
# user_input = {
# 'checked_sentence': all_news[-1].get("decsription", ""),
# 'knowledge': all_news[-1].get("decsription", ""),
# 'text': user_input_text,
# 'history': user_input_history,
# 'news_fact': "all ",
# 'news_title': all_news[-1].get("title", "")
# }
# input_batch.append(user_input)
# annotations_depths.append({})
# dial_ids.append(d_id)
except Exception as ex:
sentry_sdk.capture_exception(ex)
logger.exception(ex)
try:
raw_responses = []
if input_batch:
logger.info(f"skill sends to service: {input_batch}")
resp = requests.post(KNOWLEDGE_GROUNDING_SERVICE_URL,
json={"batch": input_batch},
timeout=1.5)
raw_responses = resp.json()
logger.info(f"skill receives from service: {raw_responses}")
else:
responses = [[""]]
confidences = [[0.0]]
attributes = [[{}]]
logger.info(
f"Collected no hypotheses, exiting with {list(zip(responses, confidences, attributes))}"
)
return jsonify(list(zip(responses, confidences, attributes)))
dial_ids = np.array(dial_ids)
attributes = []
confidences = []
responses = []
for i, dialog in enumerate(dialogs_batch):
curr_attributes = []
curr_confidences = []
curr_responses = []
for curr_i in np.where(dial_ids == i)[0]:
attr = {
"knowledge_paragraph":
input_batch[curr_i]["knowledge"],
"knowledge_checked_sentence":
input_batch[curr_i]["checked_sentence"],
"can_continue":
CAN_NOT_CONTINUE,
"confidence_case":
"",
}
already_was_active, short_long_response = get_penalties(
dialog["bot_utterances"], raw_responses[curr_i])
curr_nounphrase_search = nounphrases[i].search(
raw_responses[curr_i]) if nounphrases[i] else False
curr_entities_search = entities[i].search(
raw_responses[curr_i]) if entities[i] else False
no_penalties = False
fact_random_penalty = 0.0
topic = chosen_topics.get(i, "")
chosen_topic_fact_flag = input_batch[curr_i].get(
"chosen_topic_fact", "")
curr_news_fact = input_batch[curr_i].get("news_fact", "")
add_intro = ""
if topic and chosen_topic_fact_flag:
add_intro = f"Okay, Let's chat about {topic}. "
confidence = HIGHEST_CONFIDENCE
no_penalties = True
attr[
"confidence_case"] += f"topic_fact: {chosen_topic_fact_flag} "
attr["response_parts"] = ["prompt"]
elif input_batch[curr_i].get("news_api_fact", ""):
add_intro = random.choice([
"Sounds like ",
"Seems like ",
"Makes sense. ",
# "Here's what I've heard: ", "Here's something else I've heard: ",
"It reminds me that",
"This comes to my mind: ",
"",
])
no_penalties = True
confidence = HIGHEST_CONFIDENCE
attr["confidence_case"] += "news_api_fact "
elif input_batch[curr_i].get("fact_random_fact", ""):
fact_random_penalty = annotations_depths[curr_i].get(
"fact_random", 0.0)
confidence = DEFAULT_CONFIDENCE
attr["confidence_case"] += "fact_random_fact "
elif curr_news_fact:
if curr_news_fact != "all":
confidence = NOUNPHRASE_ENTITY_CONFIDENCE
else:
confidence = DEFAULT_CONFIDENCE
curr_news_title = input_batch[curr_i].get(
"news_title", "")
if curr_news_title:
add_intro = f"I have just read that {curr_news_title}. "
attr["confidence_case"] += "news_fact: " + curr_news_fact
elif (curr_nounphrase_search
or curr_entities_search) and lets_chat_about_flags[i]:
confidence = HIGHEST_CONFIDENCE
attr[
"confidence_case"] += "nounphrase_entity_and_lets_chat_about "
attr["response_parts"] = ["prompt"]
elif curr_nounphrase_search or curr_entities_search:
confidence = NOUNPHRASE_ENTITY_CONFIDENCE
attr["confidence_case"] += "nounphrase_entity "
elif lets_chat_about_flags[i]:
confidence = LETS_CHAT_ABOUT_CONFIDENDENCE
attr["confidence_case"] += "lets_chat_about "
attr["response_parts"] = ["prompt"]
else:
confidence = DEFAULT_CONFIDENCE
attr["confidence_case"] += "default "
acronym_flag = ABBRS.search(raw_responses[curr_i])
if acronym_flag:
confidence = ABBRS_CONFIDENCE
attr["confidence_case"] += f"acronyms: {acronym_flag} "
logger.debug(f"KG skill: found acronyms: {acronym_flag}")
special_char_flag = special_char_re.search(
raw_responses[curr_i])
if special_char_flag:
confidence = HAS_SPEC_CHAR_CONFIDENCE
attr["confidence_case"] += "special_char "
logger.debug(
f"KG skill: found special_char: {special_char_flag}")
if special_intents_flags[i]:
confidence = 0.0
attr["confidence_case"] += "special_intents "
logger.debug("KG skill: found special_intents")
greetings_farewells_flag = greetings_farewells_re.search(
raw_responses[curr_i])
if greetings_farewells_flag:
confidence = 0.0
attr["confidence_case"] += "greetings_farewells "
logger.debug(
f"KG skill: found greetings_farewells: {greetings_farewells_flag}"
)
penalties = (
annotations_depths[curr_i].get("retrieved_fact", 0.0) +
fact_random_penalty + already_was_active +
short_long_response if not no_penalties else 0.0)
confidence -= penalties
if any([
acronym_flag,
special_char_flag,
special_intents_flags[i],
greetings_farewells_flag,
short_long_response,
]):
logger.debug(
f"KG skill: found penalties in response: {raw_responses[curr_i]}, skipping it"
)
continue
else:
curr_attributes.append(attr)
curr_confidences.append(max(0.0, confidence))
curr_responses.append(
re.sub(r'\s([?.!",;:](?:\s|$))', r"\1",
add_intro + raw_responses[curr_i]).replace(
" ' t", "'t"))
attributes.append(curr_attributes)
confidences.append(curr_confidences)
responses.append(curr_responses)
except Exception as ex:
sentry_sdk.capture_exception(ex)
logger.exception(ex)
responses = [[""]]
confidences = [[0.0]]
attributes = [[{}]]
logger.info(
f"knowledge_grounding_skill exec time: {time.time() - st_time}")
return jsonify(list(zip(responses, confidences, attributes)))
async def send(self, payload: Dict, callback: Callable):
try:
st_time = time.time()
dialog = deepcopy(payload["payload"]["dialogs"][0])
is_sensitive_case = is_sensitive_situation(dialog["human_utterances"][-1])
all_prev_active_skills = payload["payload"]["all_prev_active_skills"][0]
curr_topics = get_topics(dialog["human_utterances"][-1], which="cobot_topics")
curr_nounphrases = get_entities(dialog["human_utterances"][-1], only_named=False, with_labels=False)
if len(curr_topics) == 0:
curr_topics = ["Phatic"]
logger.info(f"Found topics: {curr_topics}")
for i in range(len(curr_nounphrases)):
np = re.sub(np_remove_expr, "", curr_nounphrases[i])
np = re.sub(rm_spaces_expr, " ", np)
if re.search(np_ignore_expr, np):
curr_nounphrases[i] = ""
else:
curr_nounphrases[i] = np.strip()
curr_nounphrases = [np for np in curr_nounphrases if len(np) > 0]
logger.info(f"Found nounphrases: {curr_nounphrases}")
cands = []
confs = []
human_attrs = []
bot_attrs = []
attrs = []
cands += [choice(donotknow_answers)]
confs += [0.5]
attrs += [{"type": "dummy"}]
human_attrs += [{}]
bot_attrs += [{}]
if len(dialog["utterances"]) > 14 and not is_sensitive_case:
questions_same_nps = []
for i, nphrase in enumerate(curr_nounphrases):
for q_id in NP_QUESTIONS.get(nphrase, []):
questions_same_nps += [QUESTIONS_MAP[str(q_id)]]
if len(questions_same_nps) > 0:
logger.info("Found special nounphrases for questions. Return question with the same nounphrase.")
cands += [choice(questions_same_nps)]
confs += [0.5]
attrs += [{"type": "nounphrase_question"}]
human_attrs += [{}]
bot_attrs += [{}]
link_to_question, human_attr = get_link_to_question(dialog, all_prev_active_skills)
if link_to_question:
_prev_bot_uttr = dialog["bot_utterances"][-2]["text"] if len(dialog["bot_utterances"]) > 1 else ""
_bot_uttr = dialog["bot_utterances"][-1]["text"] if len(dialog["bot_utterances"]) > 0 else ""
_prev_active_skill = (
dialog["bot_utterances"][-1]["active_skill"] if len(dialog["bot_utterances"]) > 0 else ""
)
_no_to_first_linkto = any([phrase in _bot_uttr for phrase in LINK_TO_PHRASES])
_no_to_first_linkto = _no_to_first_linkto and all(
[phrase not in _prev_bot_uttr for phrase in LINK_TO_PHRASES]
)
_no_to_first_linkto = _no_to_first_linkto and is_no(dialog["human_utterances"][-1])
_no_to_first_linkto = _no_to_first_linkto and _prev_active_skill != "dff_friendship_skill"
_if_switch_topic = is_switch_topic(dialog["human_utterances"][-1])
bot_uttr_dict = dialog["bot_utterances"][-1] if len(dialog["bot_utterances"]) > 0 else {}
_if_choose_topic = if_choose_topic(dialog["human_utterances"][-1], bot_uttr_dict)
_is_ask_me_something = ASK_ME_QUESTION_PATTERN.search(dialog["human_utterances"][-1]["text"])
if len(dialog["human_utterances"]) > 1:
_was_cant_do = "cant_do" in get_intents(dialog["human_utterances"][-2]) and (
len(curr_nounphrases) == 0 or is_yes(dialog["human_utterances"][-1])
)
_was_cant_do_stop_it = "cant_do" in get_intents(dialog["human_utterances"][-2]) and is_no(
dialog["human_utterances"][-1]
)
else:
_was_cant_do = False
_was_cant_do_stop_it = False
if _was_cant_do_stop_it:
link_to_question = "Sorry, bye! #+#exit"
confs += [1.0] # finish dialog request
elif _no_to_first_linkto:
confs += [0.99]
elif _is_ask_me_something or _if_switch_topic or _was_cant_do or _if_choose_topic:
confs += [1.0] # Use it only as response selector retrieve skill output modifier
else:
confs += [0.05] # Use it only as response selector retrieve skill output modifier
cands += [link_to_question]
attrs += [{"type": "link_to_for_response_selector"}]
human_attrs += [human_attr]
bot_attrs += [{}]
facts_same_nps = []
for i, nphrase in enumerate(curr_nounphrases):
for fact_id in NP_FACTS.get(nphrase, []):
facts_same_nps += [
f"Well, now that you've mentioned {nphrase}, I've remembered this. {FACTS_MAP[str(fact_id)]}. "
f"{(opinion_request_question() if random.random() < ASK_QUESTION_PROB else '')}"
]
if len(facts_same_nps) > 0 and not is_sensitive_case:
logger.info("Found special nounphrases for facts. Return fact with the same nounphrase.")
cands += [choice(facts_same_nps)]
confs += [0.5]
attrs += [{"type": "nounphrase_fact"}]
human_attrs += [{}]
bot_attrs += [{}]
total_time = time.time() - st_time
logger.info(f"dummy_skill exec time: {total_time:.3f}s")
asyncio.create_task(
callback(task_id=payload["task_id"], response=[cands, confs, human_attrs, bot_attrs, attrs])
)
except Exception as e:
logger.exception(e)
sentry_sdk.capture_exception(e)
asyncio.create_task(callback(task_id=payload["task_id"], response=e))
def pickup_topic_and_start_small_talk(dialog):
"""
Pick up topic for small talk and return first response.
Args:
dialog: dialog from agent
Returns:
Tuple of (response, topic, confidence)
"""
last_user_uttr = dialog["human_utterances"][-1]
if len(dialog["bot_utterances"]) > 0:
last_bot_uttr = dialog["bot_utterances"][-1]
else:
last_bot_uttr = {"text": "---", "annotations": {}}
topic_user_wants_to_discuss = which_topic_lets_chat_about(
last_user_uttr, last_bot_uttr)
if if_choose_topic(last_user_uttr, last_bot_uttr) or if_switch_topic(
last_user_uttr["text"].lower()):
# user asks bot to chose topic: `pick up topic/what do you want to talk about/would you like to switch topic`
# or bot asks user to chose topic and user says `nothing/anything/don't know`
# if user asks to switch the topic
topic = offer_topic(dialog)
if topic in TOPIC_PATTERNS:
if topic == "me":
response = f"Let's talk about you. " + TOPIC_SCRIPTS.get(
topic, [""])[0]
elif topic == "you":
response = f"Let's talk about me. " + TOPIC_SCRIPTS.get(
topic, [""])[0]
else:
response = f"Let's talk about {topic}. " + TOPIC_SCRIPTS.get(
topic, [""])[0]
confidence = BOT_TOPIC_START_CONFIDENCE
else:
response = ""
confidence = 0.0
logger.info(f"Bot initiates script on topic: `{topic}`.")
elif topic_user_wants_to_discuss:
# user said `let's talk about [topic]` or
# bot said `what do you want to talk about/would you like to switch the topic`,
# and user answered [topic] (not something, nothing, i don't know - in this case,
# it will be gone through previous if)
topic = topic_user_wants_to_discuss
response = TOPIC_SCRIPTS.get(topic, [""])[0]
if topic in NOT_SCRIPTED_TOPICS:
confidence = YES_CONTINUE_CONFIDENCE
else:
confidence = USER_TOPIC_START_CONFIDENCE
logger.info(f"User initiates script on topic: `{topic}`.")
else:
topic = find_topics_in_substring(
dialog["human_utterances"][-1]["text"])
topic = topic[-1] if len(topic) else ""
if len(topic) > 0:
response = TOPIC_SCRIPTS.get(topic, [""])[0]
confidence = FOUND_WORD_START_CONFIDENCE
logger.info(f"Found word in user utterance on topic: `{topic}`.")
else:
topic = ""
response = ""
confidence = 0.0
return response, topic, confidence
