tatk.nlg.template.multiwoz package

Submodules

tatk.nlg.template.multiwoz.evaluate module

Evaluate NLG models on utterances of Multiwoz test dataset Metric: dataset level BLEU-4, slot error rate Usage: python evaluate.py [usr|sys|all]

tatk.nlg.template.multiwoz.evaluate.get_bleu4(dialog_acts, golden_utts, gen_utts)

tatk.nlg.template.multiwoz.nlg module

class tatk.nlg.template.multiwoz.nlg.TemplateNLG(is_user, mode='manual')

Bases: tatk.nlg.nlg.NLG

__init__(is_user, mode='manual')

Args:

is_user:

if dialog_act from user or system

mode:

• auto: templates extracted from data without manual modification, may have no match;

• manual: templates with manual modification, sometimes verbose;

• auto_manual: use auto templates first. When fails, use manual templates.

both template are dict, *_template[dialog_act][slot] is a list of templates.

generate(dialog_acts)

NLG for Multiwoz dataset

Args:

dialog_acts:

{da1:[[slot1,value1],…], da2:…}

Returns:

generated sentence

tatk.nlg.template.multiwoz.nlg.example()

tatk.nlg.template.multiwoz.nlg.read_json(filename)