JERRYLSU

【NLP】RASA Featurizer

  • 训练集stories如何构建状态state作为训练输入数据？

  • 构建的状态state作为输入X如何编码？

  • 输出y是什么？如何编码？

https://rasa.com/docs/rasa/api/core-featurization/

Dialogue Transformers

LabelTokenizerSingleStateFeaturizer creates a vector based on the feature label: All active feature labels (e.g. prev_action_listen) are split into tokens and represented as a bag-of-words. For example, actions utter_explain_details_hotel and utter_explain_details_restaurant will have 3 features in common, and differ by a single feature indicating a domain.

Labels for user inputs (intents, entities) and bot actions are featurized separately. Each label in the two categories is tokenized on a special character split_symbol (e.g. action_search_restaurant = {action, search, restaurant}), creating two vocabularies. A bag-of-words representation is then created for each label using the appropriate vocabulary. The slots are featurized as binary vectors, indicating their presence or absence at each step of the dialogue.

Transformer policy是对dialogue turn level编码，而不是token level编码。

we apply self-attention at the discourse level, attending over the sequence of dialogue turns rather than the sequence of tokens in a single turn.

一条训练数据是，X个dialogue turns和Y个action_names。

1.1 输入特征X编码

1.1.1 用户输入和系统动作词典构建

prepare_from_domain

Creates internal vocabularies for user intents and bot actions to use for featurization

user_labels = []
slot_labels = []
bot_labels = []

bot_vocab = None
user_vocab = None

只有slot_labels没有生成字典，使用列表[‘slot_s_answer_error_0’…]，也没有split(‘_’)

bot_labels = ['action_listen', 'action_restart', 'utter_ask_is_staff', 'utter_ask_visitor_reserve']
distinct_tokens = set([token for label in bot_labels for token in label.split('_')])
bot_vocab = {token: idx for idx, token in enumerate(sorted(distinct_tokens))}
bot_vocab

{'action': 0,
 'ask': 1,
 'is': 2,
 'listen': 3,
 'reserve': 4,
 'restart': 5,
 'staff': 6,
 'utter': 7,
 'visitor': 8}

user_labels = ['intent_ask_human_service', 'intent_bye', 'intent_chitchat', 'intent_confirm']
distinct_tokens = set([token for label in user_labels for token in label.split('_')])
user_vocab = {token: idx for idx, token in enumerate(sorted(distinct_tokens))}
user_vocab

{'ask': 0,
 'bye': 1,
 'chitchat': 2,
 'confirm': 3,
 'human': 4,
 'intent': 5,
 'service': 6}

slot_labels = ['slot_s_answer_error_0', 'slot_s_chitchat_turn_0', 'slot_s_digits_key_0', 'slot_s_host_name_0', 'slot_s_is_call_human_0', 'slot_s_is_call_human_1', 'slot_s_is_chitchat_0', 'slot_s_is_chitchat_1', 'slot_s_is_reserve_visitor_0', 'slot_s_is_reserve_visitor_1', 'slot_s_is_same_name_0', 'slot_s_is_same_name_1', 'slot_s_is_staff_0', 'slot_s_is_staff_1', 'slot_s_is_valid_info_0', 'slot_s_is_valid_info_1', 'slot_s_is_valid_reserve_0', 'slot_s_is_valid_reserve_1', 'slot_s_is_valid_staff_0', 'slot_s_is_valid_staff_1', 'slot_s_is_visitor_0', 'slot_s_is_visitor_1', 'slot_s_phone_number_0', 'slot_s_self_name_0', 'slot_s_staff_homophonic_name_0', 'slot_s_title_name_0', 'slot_s_visitor_homophonic_name_0']

1.1.2 encode——词袋模型

总特征向量维度：num_features = 用户输入（意图和实体）字典维度 + 槽维度 + 系统action字典维度

user_feature_len = len(user_vocab)
print("user feature len: {}".format(user_feature_len))
slot_feature_len = len(slot_labels)
print("slot feature len: {}".format(slot_feature_len))
bot_feature_len = len(bot_vocab)
print("bot feature len: {}".format(bot_feature_len))
num_features = len(user_vocab) + len(slot_labels) + len(bot_vocab)
print("num_feature = user vocab + slot labels + bot vocab: {}".format(num_features))

user feature len: 7
slot feature len: 27
bot feature len: 9
num_feature = user vocab + slot labels + bot vocab: 43

# 特征化向量，固定长度。

import numpy as np
used_features = np.zeros(num_features, dtype=float)
used_features 

array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
       0., 0., 0., 0., 0., 0.])

# 通过if判断状态属于用户、槽、还是系统动作，以确定在向量num_features中的偏移量offset。
# 用户的意图实体 + 槽 + 系统动作

# 输入的状态state（包括用户输入的意图实体、槽、系统动作）
state = {'entity_e_phone_number': 1.0,
         'intent_deny+inform_name_self+inform_phone_number_self': 1.0, 
         'prev_action_listen': 1.0, 'entity_e_name': 1.0, 
         'entity_e_four_digits': 1.0}

def split_state_name(state_name: str):
    """Split multiple intents with '+' and '_'.
        Add the string of 'intent'.
    e.g.
    "intent_deny+inform_name_self+inform_phone_number_self" ->
    ['intent', 'deny', 'inform', 'name', 'self', 'inform', 'phone', 'number', 'self', 'intent', 'intent']
    """
    intents = state_name.split('+')
    intents_num, res = len(intents), []
    for words in intents:
        res.extend(words.split('_'))
    for _ in range(intents_num - 1):
        res.append('intent')
    return res

used_features = np.zeros(num_features, dtype=float)
idx = 0
PREV_PREFIX = "prev_"
for state_name, prob in state.items():
    idx += 1
    print()
    print('state name {}: {}'.format(idx, state_name))

    if state_name in user_labels:                            # 用户输入编码入向量used_features
        #for t in [word for words in state_name.split('+') for word in words.split('_')]: # Bingo!!!
        for t in split_state_name(state_name):
        #for t in state_name.split('_'):
            idx = user_vocab[t]
            used_features[idx] += prob
            print(t)
            print(used_features)

    elif state_name in slot_labels:                          # 槽编码入向量used_features
        offset = len(user_vocab)
        idx = slot_labels.index(state_name)
        used_features[offset + idx] += prob

    elif state_name[len(PREV_PREFIX) : ] in bot_labels:      # 系统动作编码入向量used_features
        action_name = state_name[len(PREV_PREFIX) :]
        for t in action_name.split('_'):
            offset = len(user_vocab) + len(slot_labels)
            idx = bot_vocab[t]
            used_features[offset + idx] += prob
            print(t)
            print(used_features)

state name 1: entity_e_phone_number
entity
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
e
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
phone
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
number
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 1. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]

state name 30: intent_deny+inform_name_self+inform_phone_number_self
intent
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 0. 1. 0. 0. 0. 1. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
deny
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 0. 1. 0. 0. 0. 1. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
inform
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 1. 1. 0. 0. 0. 1. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
name
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 1. 1. 0. 0. 1. 1. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
self
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 1. 1. 0. 0. 1. 1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
inform
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 1. 0. 0. 1. 1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
phone
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 1. 0. 0. 1. 1. 0. 0. 2. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
number
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 1. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
self
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 1. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
intent
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 2. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
intent
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]

state name 26: prev_action_listen
action
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
listen
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 1. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]

state name 33: entity_e_name
entity
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 1. 2. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
e
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 2. 2. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 1. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
name
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 2. 2. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 2. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]

state name 29: entity_e_four_digits
entity
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 2. 3. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 2. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
e
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 3. 3. 0. 0. 0. 0. 0.
 2. 3. 0. 0. 2. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
four
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 3. 3. 1. 0. 0. 0. 0.
 2. 3. 0. 0. 2. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]
digits
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 1. 3. 3. 1. 0. 0. 0. 0.
 2. 3. 0. 0. 2. 2. 0. 0. 2. 0. 0. 0. 2. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0.]

state_nam = 'intent_deny+inform_name_self'

def split_state_name(state_name: str):
    intents = state_name.split('+')
    intents_num = len(intents)
    res = []
    for words in intents:
        res.extend(words.split('_'))
    for _ in range(intents_num - 1):
        res.append('intent')
    return res

split_state_name(state_nam)

['intent', 'deny', 'inform', 'name', 'self', 'intent']

[word for words in state_name.split('+') for word in words.split('_')]

['entity', 'e', 'four', 'digits']

if PREV_PREFIX + ACTION_LISTEN_NAME in state:前一轮是action_listen,才对用户的输入（即，意图和实体）编码。而后，几轮虽然包含前几轮信息，但对信息中的用户输入不再编码

  • X的shape=（1，4，153）1th维是story数，2维是对话轮数，3维是每轮的编码vector长度。固定长度，有padding

1.2 输出Y编码

  1. 对action进行one-hot编码。 ‘action_listen’ —-> ‘1,0,0,…,0,0,0’ (62,)维向量

  2. 构造bot_action字典，”“拆分。 ‘action_listen’ —-> ‘1, 0, 0, .., 1, …, 0, 0’ (62, 72)维矩阵。62个actions，每个action用72维表示，72是’‘拆分所有action_names后构成的字典数。

  3. one-hot向量转bot_action字典向量。np.argmax(onehot vector)获取索引, 查(62, 72)矩阵。

Y 最终的输出维度是(1dim, 2dim, 3dim) = (1, 4, 72): 1 story, 4 turns dialogue, 72 vector of action dict。对于FullDialogueTrackerFeaturizer,第2个维度是对话轮数大小，按照最长story长度padding -1, 用-1补齐的轮对应的标签是action_listen。

FullDialogueTrackerFeaturizer 和 MaxHistoryTrackerFeaturizer 的区别就是再对2nd维度上的控制，

        if y.ndim == 3 and isinstance(self, MaxHistoryTrackerFeaturizer):
            # if it is MaxHistoryFeaturizer, remove time axis
            y = y[:, 0, :]

## Generated Story case-0-56898
* inform_face_name{"e_face_name_1": "张春梅"}
    - action_set_face_name
    - slot{"s_face_name_1": "张春梅"}
    - utter_goodbye

# domain

action_names = ['action_listen', 'action_restart', 'action_session_start', 'action_default_fallback', 'action_deactivate_form', 'action_revert_fallback_events', 'action_default_ask_affirmation', 'action_default_ask_rephrase', 'action_back', 'action_confirm_homophonic_name', 'action_judge_same_name_order_or_department_validity', 'action_match_answer_info', 'action_match_reserve_visitor_name', 'action_match_staff_name', 'action_set_face_name', 'action_set_full_name', 'action_set_host_name', 'action_set_reality_name_by_face', 'action_set_reality_name_by_joint', 'action_set_same_name_order_or_department', 'action_set_self_name', 'action_set_user_domain', 'action_set_user_info', 'utter_ask_is_staff', 'utter_ask_staff_digits_key', 'utter_ask_staff_digits_key_again', 'utter_ask_staff_name', 'utter_ask_staff_name_and_digits_key', 'utter_ask_user_is_staff', 'utter_ask_visitor_host', 'utter_ask_visitor_host_full_name', 'utter_ask_visitor_name', 'utter_ask_visitor_name_and_phone', 'utter_ask_visitor_phone', 'utter_ask_visitor_phone_again', 'utter_ask_visitor_phone_error', 'utter_ask_visitor_phone_error_again', 'utter_ask_visitor_phone_no_pass', 'utter_ask_visitor_reserve', 'utter_call_human_service', 'utter_chitchat', 'utter_chitchat_again', 'utter_chitchat_once_again', 'utter_confirm_host_name', 'utter_confirm_staff_name', 'utter_confirm_visitor_name', 'utter_correct_answer', 'utter_default', 'utter_error_authentication', 'utter_error_catch', 'utter_goodbye', 'utter_greet_short', 'utter_help_find_people', 'utter_staff_digits_key_error', 'utter_staff_digits_key_error_again', 'utter_staff_inform_no_collection', 'utter_staff_invalid_name', 'utter_staff_welcome', 'utter_visitor_lack_host_information', 'utter_visitor_lack_reserve_name', 'utter_visitor_wait', 'utter_wrong_answer']
bot_vocab = {'action': 0, 'affirmation': 1, 'again': 2, 'and': 3, 'answer': 4, 'ask': 5, 'authentication': 6, 'back': 7, 'by': 8, 'call': 9, 'catch': 10, 'chitchat': 11, 'collection': 12, 'confirm': 13, 'correct': 14, 'deactivate': 15, 'default': 16, 'department': 17, 'digits': 18, 'domain': 19, 'error': 20, 'events': 21, 'face': 22, 'fallback': 23, 'find': 24, 'form': 25, 'full': 26, 'goodbye': 27, 'greet': 28, 'help': 29, 'homophonic': 30, 'host': 31, 'human': 32, 'info': 33, 'inform': 34, 'information': 35, 'invalid': 36, 'is': 37, 'joint': 38, 'judge': 39, 'key': 40, 'lack': 41, 'listen': 42, 'match': 43, 'name': 44, 'no': 45, 'once': 46, 'or': 47, 'order': 48, 'pass': 49, 'people': 50, 'phone': 51, 'reality': 52, 'rephrase': 53, 'reserve': 54, 'restart': 55, 'revert': 56, 'same': 57, 'self': 58, 'service': 59, 'session': 60, 'set': 61, 'short': 62, 'staff': 63, 'start': 64, 'user': 65, 'utter': 66, 'validity': 67, 'visitor': 68, 'wait': 69, 'welcome': 70, 'wrong': 71}
num_actions = len(action_names)
num_actions

62

trackers_as_actions = ['action_listen', 'action_set_face_name', 'utter_goodbye', 'action_listen']

['action_listen', 'action_set_face_name', 'utter_goodbye', 'action_listen']

# Encode system action as one-hot vector.

labels = [] # Multi_story

def action_as_one_hot(action):
    y = np.zeros(num_actions, dtype=int)
    index_for_action = action_names.index(action)
    #print(f'Index for action: {index_for_action}')
    y[index_for_action] = 1
    return y
    #print(f'One-hot of y: {y}\n y shape: {y.shape}')

story_labels = [action_as_one_hot(action) for action in trackers_as_actions] # one story and multi_turns
labels.append(story_labels)
y = np.array(labels)
y.shape # (1dim, 2dim, 3dim) = (1, 4, 62) 1 story, 4 turns dialogue, 62 one-hot vector of action

(1, 4, 62)

# Create matrix with all actions from domain encoded in rows as bag of words

def create_encoded_all_actions() -> np.ndarray:
    encoded_all_actions = np.zeros((num_actions, len(bot_vocab)), dtype=np.int32)
    for idx, name in enumerate(action_names):
        for t in name.split('_'):
            encoded_all_actions[idx, bot_vocab[t]] = 1
    return encoded_all_actions

encoded_all_label_ids = create_encoded_all_actions()
encoded_all_label_ids.shape # 62 num of action, 72 bot vocab size

(62, 72)

# extract actual training data to feed to tf session

print(y.shape)
y # One-hot encode. (1, 4, 62): 1 story, 4 turns dialogue(one turn is one action), 62 vector of action

(1, 4, 62)





array([[[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]])

label_ids = y.argmax(-1)
label_ids

array([[ 0, 14, 50,  0]])

d = [-1,-1,-1]
d = np.array(d)
d.argmax(-1)

0

label_ids.shape

(1, 4)

# full dialogue
res = []
for seq_label_ids in label_ids:
    for label_idx in seq_label_ids:
        res.append(encoded_all_label_ids[label_idx])
res

[array([1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0], dtype=int32),
 array([1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0], dtype=int32),
 array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        1, 0, 0, 0, 0, 0], dtype=int32),
 array([1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0], dtype=int32)]

res = np.stack(res)
res

array([[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0],
       [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        1, 0, 0, 0, 0, 0],
       [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0]], dtype=int32)

label_ids.shape # explicitly add last dimension to label_ids to track correctly dynamic sequences

(1, 4)

label_ids = np.expand_dims(label_ids, -1)
label_ids

array([[[ 0],
        [14],
        [50],
        [ 0]]])

FullDialogueTrackerFeaturizer

# Training data is padded up to the length of the longest dialogue with -1.

!jupyter nbconvert --to markdown featurizer.ipynb

[NbConvertApp] Converting notebook featurizer.ipynb to markdown
[NbConvertApp] Writing 27001 bytes to featurizer.md

Share on: Twitter ❄ Facebook ❄ Email

Comments

Related Posts

Reading Time

~21 min read

Published

Category

NLP

Tags

Contact