verify the existence of the data file

make agent action prediction according to the threshold Inputs: act_probs: shape = (sample_nb, act_vocab_size), predicted action

return a list of lists, [[(w1, tag1), (w2, tag2), (w3, tag3)], [...], [...]]

Inputs: intent_probs: shape = (sample_nb, intent_vocab_size), predicted probs for intent prediction intent_trueLabel: shape =

make decision of label prediction given specific threshold

padding (window_size -1) zero vectors in front of vecs, and chunking padded vecs iteratively using a sliding window. The chunking ra

calculate frame-level accuracy = hit / sample_nb inputs: pred: shape = (sample_nb, dim_size), predicted ids matrix

Evaluation for slot tagging. Inputs: tag_probs: shape = (sample_nb, maxlen_userUtter, tag_vocab_size), predicted probs

get token2id, id2token, vocab_size for intent, act, tag, or word. need to reserve 1 for '<unk>', and reserve 0 for '<pad>'. note: th

calculate BoW feature Input: an 2-darray user utterance with zero padding Output: 2-darray

transform into a 1hot matrix Input: y_seq: shape = (sample_nb, maxlen_userUtter), elements are token ids. nb_classes:

convert into binary vectors. Inputs: intents: shape = (sample_nb,) vocab_size: scalar, vocabulary size in

write user utter and user tags into a file. each line constrains with the format: w1 w2 w3\ttag1 tag2 tag3 Inputs: sents

convert textual utter and user tags into a list of lists that contain lists of (w, t) pairs

encode utterance or slot tags into id sequence. 0s for padding, 1s for unk. return a matrix with shape=(sample_nb, maxlen) e.

[Q] how to deal with the case: the tag, intent, act in test or dev set do not exist in train set, but they are predefined

calculate performance score Input: y_true: true tags, [[(u'w1', tag1), (u'w2', tag2), (u'w3', tag3)], [...], [...]] y

userTags_pred: [[(w1, tag1), (w2, tag2)],[]]

train intent prediction model using human annotated data Input: bag-of-words of user utterances, Output: indicator matrix of agent actions

train slot tagging model using human annotated data Input: userUtter Output: target userTags

construct a search dictionary: {sudialogID_utterIndex: normalized_utterance}

Plot learning curves of losses, and save it as .png file. Input: loss_dct: {'train': [], 'val': [], 'train_tagging':[], 'tr