def main(in_dataset):
# Hough and Schlangen 2015 config
disf = DeepDisfluencyTagger(
config_file=
"deep_disfluency/deep_disfluency/experiments/experiment_configs.csv",
config_number=21,
saved_model_dir=
"deep_disfluency/deep_disfluency/experiments/021/epoch_40")
for idx, row in in_dataset.iterrows():
print ' '.join(row['utterance'])
tagger_input = [(word, pos, None)
for word, pos in zip(row['utterance'], row['pos'])]
import pdb
pdb.set_trace()
print ' '.join(disf.tag_utterance(tagger_input))
def tag_words(txt_file):
disf = DeepDisfluencyTagger(
config_file="../deep_disfluency/experiments/experiment_configs.csv",
config_number=21,
saved_model_dir="../deep_disfluency/experiments/021/epoch_40")
with open(txt_file, "r") as f:
for lines in f:
line = lines.split(".")
for words in line:
word = words.split()
for i in word:
disf.tag_new_word(i)
tags = disf.output_tags
return tags
def main(in_dataset, in_mode):
# Hough and Schlangen 2015 config
disf = DeepDisfluencyTagger(
config_file=
"deep_disfluency/deep_disfluency/experiments/experiment_configs.csv",
config_number=21,
saved_model_dir=
"deep_disfluency/deep_disfluency/experiments/021/epoch_40")
for key, value in eval_babi(disf, in_dataset).iteritems():
print '{}:\t{:.3f}'.format(key, value)
from deep_disfluency.tagger.deep_tagger import DeepDisfluencyTagger
# Initialize the tagger from the config file with a config number
# and saved model directory
MESSAGE = """1. Disfluency tagging on pre-segmented utterances
tags repair structure incrementally and other edit terms <e/>
(Hough and Schlangen Interspeech 2015 with an RNN)
"""
print MESSAGE
disf = DeepDisfluencyTagger(config_file="experiments/experiment_configs.csv",
config_number=21,
saved_model_dir="experiments/021/epoch_40")
# Tag each word incrementally
# Notice the incremental diff
# Set diff_only to False if you want the whole utterance's tag each time
with_pos = False
print "tagging..."
if with_pos:
# if POS is provided use this:
print disf.tag_new_word("john", pos="NNP")
print disf.tag_new_word("likes", pos="VBP")
print disf.tag_new_word("uh", pos="UH")
print disf.tag_new_word("loves", pos="VBP")
print disf.tag_new_word("mary", pos="NNP")
else:
# else the internal POS tagger tags the words incrementally
print disf.tag_new_word("john")
print disf.tag_new_word("likes")
print disf.tag_new_word("uh")
print disf.tag_new_word("loves")
word_graph[
new_output_start:],
self.disf.get_output_tags(with_words=False)[
new_output_start:],
range(
new_output_start, len(word_graph))
):
print w, h, i
if ("<e" in h or "<rps" in h) and \
i not in self.words_with_disfluency:
self.label.config(text=str(self.counter.next()))
self.master.update()
self.words_with_disfluency.append(i)
except:
print "FAILED TO UPDATE"
self.master.update()
self.asr = IBMWatsonASR("credentials.json",
new_word_hypotheses_handler)
self.asr.listen()
if __name__ == '__main__':
disf = DeepDisfluencyTagger(
config_file="deep_disfluency/experiments/experiment_configs.csv",
config_number=35,
saved_model_dir="deep_disfluency/experiments/035/epoch_6",
use_timing_data=True
)
root = Tk()
my_gui = DisfluencyGUI(root, disf)
print 'Finished extracting features.'
# 4. Train the model on the transcripts (and audio data if available)
# NB each of these experiments can take up to 24 hours
systems_best_epoch = {}
if train_models:
feature_matrices_filepath = THIS_DIR + '/../data/disfluency_detection/' + \
'feature_matrices/train'
validation_filepath = THIS_DIR + '/../data/disfluency_detection/' + \
'feature_matrices/heldout'
# train until convergence
# on the settings according to the numbered experiments in
# experiments/config.csv file
for exp in experiments:
disf = DeepDisfluencyTagger(config_file=THIS_DIR +
"/experiment_configs.csv",
config_number=exp)
exp_str = '%03d' % exp
e = disf.train_net(train_dialogues_filepath=feature_matrices_filepath,
validation_dialogues_filepath=validation_filepath,
model_dir=THIS_DIR + '/' + exp_str,
tag_accuracy_file_path=THIS_DIR +
'/results/tag_accuracies/{}.text'.format(exp_str))
systems_best_epoch[exp] = e
else:
# 33 RNN simple tags, disf + utt joint
# 34 RNN complex tags, disf + utt joint
# 35 LSTM simple tags, disf + utt joint
# 36 LSTM complex tags, disf + utt joint
# 37 LSTM simple tags, disf only
# 38 LSTM simple tags, utt only
# (38, 8, 'LSTM (TTO only)'),
# (39, 2, 'LSTM (complex tags)')
# 4. Test the models on the test transcripts according to the best epochs
# from training.
# The output from the models is made in the folders
# For now all use timing data
if test_models:
print "testing models..."
for exp, best_epoch in sorted(systems_best_epoch.items(), lambda x: x[0]):
exp_str = '%03d' % exp
# load the model
disf = DeepDisfluencyTagger(
config_file=THIS_DIR + '/experiment_configs.csv',
config_number=exp,
saved_model_dir=THIS_DIR +
'/{0}/epoch_{1}'.format(exp_str, best_epoch))
# simulating (or using real) ASR results
# for now just saving these in the same folder as the best epoch
# also outputs the speed
partial_string = '_partial' if partial else ''
disf.incremental_output_from_file(
THIS_DIR + '/../data/disfluency_detection/switchboard/' +
'swbd_disf_heldout{}_data_timings.csv'.format(partial_string),
target_file_path=THIS_DIR +
'/{0}/epoch_{1}/'.format(exp_str, best_epoch) +
'heldout_output_increco.text')
# 5. To get the numbers run the notebook:
# experiments/analysis/EACL_2017/EACL_2017.ipynb