Optimizing Segmentation Strategies for

Optimizing Segmentation Strategies for Simultaneous Speech Translation
Yusuke Oda
Graham Neubig
Sakriani Sakti
Tomoki Toda
Satoshi Nakamura
Nara Institute of Science and Technology (NAIST), Japan
Simultaneous Speech Translation (SST)
Framework
0.20
MT
今から18分間で、
皆様を旅に
ご案内します。
SST systems translate speech into the other
language as fast as possible.
Experiments & Results
Set of
segmentation
boundaries
EV
Training
corpus
Segmentation
model
En-De
Punct-Predict (conventional)
Search
segmentation
Learn
features
Assumption: good segmentation maximizes evaluation measure:
BLEU
In the next 18 minutes,
I'm going to take you
on a journey.
Translate
Algorithm of Optimizing Segmentation
RP (conventional)
Greedy
0.15
Greedy+Grouping
En-Ja
Segmentation for SST System
SST systems require a method that splits the
input sequence at appropriate positions.
= Segmentation strategy
w/o segmentation
Split
Sentence-wise
evaluation
measure
today
MT
e.g.
BLEU+1, RIBES
I'm going to take you on a journey
彼
予定です
To obtain
on a journey
皆様を案内する
0.30
today he ate lunch but she left
Evaluation
Measure
0.7
今日 彼 昼食を食べたが彼女は去った
Hypothesis
今日 彼は食べた ランチ彼女は去った
0.4
today he ate lunch but she left
今日 彼は昼食を食べた だが彼女は去った
0.6
•
Using pausing of speech recognition
today he ate lunch but she left
今日 彼は昼食を食べたが 彼女は去った
0.9
[Bangalore et al., NAACL HLT 2012]
today he ate lunch but she left
今日 彼は昼食を食べたが彼女 左
0.2
Using predicted punctuation and others
[Sridhar et al., NAACL HLT 2013]
Using probability of phrase reordering
[Fujita et al., InterSpeech 2013]
Problem … All heuristic approaches
• No guarantee of good MT accuracy
Our New Methods
Explicitly construct the optimal strategies for MT.
→ Assuming segmentation as an
optimization problem given MT system.
Already chosen
Choose this boundary and add it to
Feature Grouping
Greedy search … over-fitting → Introduce grouping constraint.
he
ate
PRP
I
PRP
lunch
VBD
ate
VBD
Greedy method
accuracy high for
training data
but low for test data
→ Over-fitting
BLEU
, we use greedy search-based boundary selection.
today he ate lunch but she left
•
NN
an
DT
15
昼食を食べた
Conventional Methods
•
10
Greedy+Grouping splits input sequence 2-3 times
frequently while keeping same translation performance.
ate lunch
今日彼昼食を食べた
Source Sentence
旅に
5
#words/segment
Greedy Search
w/ segmentation
take you
he
今日
Join
皆様を…
I'm going to
0.10
0
today he ate lunch
apple
NN
but
CC
and
CC
she
PRP
an
DT
left
VBD
orange
NN
Boundaries with same surrounding POS must be chosen at the same time.
En-De
0.20
En-Ja
Greedy
0.10
Greedy+Grouping
0
5
10
15
#words/segment
Grouping constraint can efficiently avoid over-fitting
from greedy search.
Conclusion & Future Works
Our methods find segmentation using no heuristics
Our methods can split input sentence more frequently
while keeping translation performance.
Future works are:
• Developing algorithm scalable w.r.t. corpus size.
• Introducing multiple features.
Our implementation is available in:
http://odaemon.com/docs/codes/greedyseg.html