tatk.dst.mdbt package

Submodules

tatk.dst.mdbt.mdbt module

class tatk.dst.mdbt.mdbt.MDBT(ontology_vectors, ontology, slots, data_dir)

Bases: tatk.dst.dst.DST

A multi-domain belief tracker, adopted from https://github.com/osmanio2/multi-domain-belief-tracking.

__init__(ontology_vectors, ontology, slots, data_dir)

Initialize self. See help(type(self)) for accurate signature.

detect_requestable_slots(observation)
init_session()

Init the class variables for a new session.

normalize_history(history)

Replace zero-length history.

restore()
update(user_act=None)

Update the dialog state.

update_batch(batch_action)

tatk.dst.mdbt.mdbt_util module

class tatk.dst.mdbt.mdbt_util.GRU(W_h, U_h, M_h, W_m, U_m, label_size, reuse=None, binary_output=False)

Bases: tensorflow.python.ops.rnn_cell_impl.RNNCell

Create a Gated Recurrent unit to unroll the network through time for combining the current and previous belief states

__init__(W_h, U_h, M_h, W_m, U_m, label_size, reuse=None, binary_output=False)
property output_size

Integer or TensorShape: size of outputs produced by this cell.

property state_size

size(s) of state(s) used by this cell.

It can be represented by an Integer, a TensorShape or a tuple of Integers or TensorShapes.

tatk.dst.mdbt.mdbt_util.define_CNN_model(utter, num_filters=300, name='r')

Better code for defining the CNN model.

tatk.dst.mdbt.mdbt_util.evaluate_model(sess, model_variables, val_data, summary, batch_id, i)

Evaluate the model against validation set

Parameters

  • sess – training session

  • model_variables – all model input variables

  • val_data – validation data

  • summary – For tensorboard

  • batch_id – where we are in the training data

  • i – the index of the validation data to load

Returns

evaluation accuracy and the summary

tatk.dst.mdbt.mdbt_util.generate_batch(dialogues, batch_no, batch_size, ontology_size)

Generate examples for minibatch training

Parameters

  • dialogues – list(num of turns, user_input vectors, system_response vectors, labels)

  • batch_no – where we are in the training data

  • batch_size – number of dialogues to generate

  • ontology_size – no_slots

Returns

list(user_input, system_response, labels, user_sentence_length, system_sentence_length, number of turns)

tatk.dst.mdbt.mdbt_util.get_available_devs()
tatk.dst.mdbt.mdbt_util.get_metrics(predictions, true_predictions, no_turns, mask, num_slots)
tatk.dst.mdbt.mdbt_util.load_ontology(url, word_vectors)

Load the ontology from a file

Parameters

  • url – to the ontology

  • word_vectors – dictionary of the word embeddings [words, vector_dimension]

Returns

list([domain-slot-value]), [no_slots, vector_dimension]

tatk.dst.mdbt.mdbt_util.load_word_vectors(url)

Load the word embeddings from the url

Parameters

url – to the word vectors

Returns

dict of word and vector values

tatk.dst.mdbt.mdbt_util.load_woz_data(data, word_vectors, ontology, url=True)

Load the woz3 data and extract feature vectors

Parameters

  • data – the data to load

  • word_vectors – word embeddings

  • ontology – list of domain-slot-value

  • url – Is the data coming from a url, default true

Returns

list(num of turns, user_input vectors, system_response vectors, labels)

tatk.dst.mdbt.mdbt_util.load_woz_data_new(data, word_vectors, ontology, url=False)

Ported from load_woz_data, using tatk.util.dataloader pkg

Parameters

  • data – the data to load

  • word_vectors – word embeddings

  • ontology – list of domain-slot-value

  • url – Is the data coming from a url, default true

Returns

list(num of turns, user_input vectors, system_response vectors, labels)

tatk.dst.mdbt.mdbt_util.lstm_model(text_input, utterance_length, num_hidden, name, net_type, bidir)

Define an Lstm model that will run across the user input and system act

Parameters

  • text_input – [batch_size, max_num_turns, max_utterance_size, vector_dimension]

  • utterance_length – number words in every utterance [batch_size, max_num_turns, 1]

  • num_hidden – – int –

  • name – The name of lstm network

  • net_type – type of the network (“lstm” or “gru” or “rnn”)

  • bidir – use a bidirectional network – bool –

Returns

output at each state [batch_size, max_num_turns, max_utterance_size, num_hidden], output of the final state [batch_size, max_num_turns, num_hidden]

tatk.dst.mdbt.mdbt_util.model_definition(ontology, num_slots, slots, num_hidden=None, net_type=None, bidir=None, test=False, dev=None)

Create neural belief tracker model that is defined in my notes. It consists of encoding the user and system input, then use the ontology to decode the encoder in manner that detects if a domain-slot-value class is mentioned

Parameters

  • ontology – numpy array of the embedded vectors of the ontology [num_slots, 3*vector_dimension]

  • num_slots – number of ontology classes –int–

  • slots – indices of the values of each slot list of lists of ints

  • num_hidden – Number of hidden units or dimension of the hidden space

  • net_type – The type of the encoder network cnn, lstm, gru, rnn …etc

  • bidir – For recurrent networks should it be bidirectional

  • test – This is testing mode (no back-propagation)

  • dev – Device to run the model on (cpu or gpu)

Returns

All input variable/placeholders output metrics (precision, recall, f1-score) and trainer

tatk.dst.mdbt.mdbt_util.normalise_word_vectors(word_vectors, norm=1.0)

This method normalises the collection of word vectors provided in the word_vectors dictionary.

tatk.dst.mdbt.mdbt_util.process_booking(ontolog_term, usr_input, previous_terms)
tatk.dst.mdbt.mdbt_util.process_history(sessions, word_vectors, ontology)

Load the woz3 data and extract feature vectors

Parameters

  • data – the data to load

  • word_vectors – word embeddings

  • ontology – list of domain-slot-value

  • url – Is the data coming from a url, default true

Returns

list(num of turns, user_input vectors, system_response vectors, labels)

tatk.dst.mdbt.mdbt_util.process_text(text, word_vectors, ontology=None, print_mode=False)

Process a line/sentence converting words to feature vectors

Parameters

  • text – sentence

  • word_vectors – word embeddings

  • ontology – The ontology to do exact matching

  • print_mode – Log the cases where the word is not in the pre-trained word vectors

Returns

[length of sentence, 300]

tatk.dst.mdbt.mdbt_util.process_turn(turn, word_vectors, ontology)

Process a single turn extracting and processing user text, system response and labels

Parameters

  • turn – dict

  • word_vectors – word embeddings

  • ontology – list(domain-slot-value)

Returns

([utterance length, 300], [utterance length, 300], [no_slots])

tatk.dst.mdbt.mdbt_util.track_dialogue(data, ontology, predictions, y)
tatk.dst.mdbt.mdbt_util.xavier_vector(word, D=300)

Returns a D-dimensional vector for the word.

We hash the word to always get the same vector for the given word.