Forest Driven Rulebased Spoken Language Translation

1
Forest Driven Rule-based Spoken Language
TranslationKAMATANI Satoshi, FURIHATA Kentaro
and CHINO TetsuroCorporate Research
Development CenterToshiba Corporation1, Komukai
Toshiba-cho, Saiwai-ku, Kawasaki, 212-8583,
Japane-mail satoshi.kamatani_at_toshiba.co.jp
CJNLP 2006The 6th China-Japan Natural Language
Processing Joint Research Promotion Conference

• 1. Introduction
• Towards realization of Spoken language
  translation (SLT)
• Major machine translation methods
• Rule based machine translation (RBMT)
• Example based machine translation (EBMT)
• Stochastic machine translation (SMT)
• SLT based on RBMT
• Market-proven translation quality
• Comparatively easy to respond to a request for
  modifications
• Beneficial use of Translation rules accumulated
  many years

2.2.4 Forest dependency analysis

• Co-occurrence Probability Model
• learned by EM-algorithm based method proposed by
  Torisawa et.al 2001
• h Head word ( modified word )
• m Modifier
• rel Relation between the head word and the
  modifier
• C Set of hidden co-occurrence classes
• Co-occurrence Preference
• INT(x) a function to round down to integer
• INT ( average of whole co-occurrence
  probabilities ) 12
• EPS

Corpus

• (???,?,????) 2
• (?????,?, ????) 1
• (?????,?,??) 3

Learning
Co-occurrence model
Co-occurrence Preference
Round down
Co-occurrence preference mj modifies mp
1.0
confliction with clause dependency
condition -0.8
But it is necessary to analyze robustly and
bridge between spoken and written language.

• 2. Forest Driven Spoken Language Translation
• 2.1 Overview
• It realizes a fair evaluation for all the
  possible candidates of interpretation without
  unreasonable suppositions
• All of syntactic/semantic preferences are
  evaluated by one integrated mechanism on the
  forest structures.

minami_B
cm_ga
Source Language Input
The enumeration of all possible interpretation
candidates.
(a)
Robust GLR Parsing
mi
mj
mp
mo
mq
mk
syntax forest
minami_B
minami_C
cm_ga
The interpretation of the spoken language
specific structure and a transfer to equivalent
but simple structure.
Pattern 0.poscase maker 0.surface?
-1.posnoun Clause type cm_ga Activate in
modifier Dependency co_occured with other
cm_ga -0.9 co_occured with ad_ha
-0.5
(b)
Syntactic Pruning
(c)
Partial forest transfer
(d)
Forest driven dependency analysis
The estimation of the optimum interpretation.
optimum dependency structure
Pattern 0.posconjunctive maker
0.surface?? -1.posverb or
-1.posauxiliary verb Clause type minami_B
(classified into class B by Minami) Active in
modified Dependency co_occured with other
cm_ga -0.8 co_occured with minami_A
-1.0
Practical use of rich translation knowledge.
(e)
Transfer Generation
Target Language Output
6
4
4
3
2
1

• 2.2 Detail of each step
• 2.2.1 Robust GLR Parsing
• Robust CFG
• Designed for fragmental inputs
• Capture clause structure
• Introduce ill-formednes marker weak
• The GLR parser leads all possible interpretations
  as a syntax forest.

????
????
????
??????
?
I
the
whose
is
book
restaurant
soup
good

a
books
restaurant
Soup
good

• 3. Evaluation
• 3.1 Conditions of Experiment
• We developed a prototype system
• Evaluation target
• Japanese to English speech translation ( Travel
  domain, Open data )
• 702 utterances ( 13.4 letters, 7.2 morphemes in
  average )
• 3.2 Parsing Accuracy
• All of the test utterances are accepted by our
  grammar.
• Reduce the number of parsing Reduce
  ambiguities
• Improve the ratio of correct dependency (CDR) in
  the forest
• T-test assessed a significant improvement with
  difference 0.05.

Example

• 2.2.2 Syntactic Pruning
• Evaluate syntactic well-formedness and prune ill
  structures
• Estimate ill-formedness for all interpretations
  in the forest.
• Assign a penalty to weak marked vertices.
• Aggregate the possible minimum penalty by
  bottom up manner.
• Extract syntactically preferred part.
• Select vertices with the lowest penalty
  stepwise top down manner.
• Extract a well-formed structure consists of
  only preferred vertices.

F a syntax forest t a syntax tree in F NF
number of t?F Dc a number of correct
dependency Dm number of dependencies in t
matched to the correct one.

• 3.3 Translation Quality
• Visual evaluation by 3 persons, majority
  decision
• Evaluate whether the translation conveys
  intention of the original utterance or not.
• Evaluate two systems
• Forest Driven SLT system
• A commercial MT Engine for written language.

• 2.2.3 Partial Forest Transfer
• Transfer transfers a tree to a tree.
• It is widely and successfully utilized in
  conventional MT.
• It converts the internal expressions based on
  the transfer rules.
• But, It can treat only one tree structure.
• Partial Forest Transfer transfers a forest to
  a forest directly.
• It bridges the gap between spoken and written
  languages.
• It reduces ambiguity in parses.

They share the same set of transfer rules!

• 4. Conclusion
• Forest driven spoken language translation system
• Robust and efficient processing for Japanese
  utterances.
• Beneficial use of the conventional rule base
  machine translator.
• Future works
• Utilization of machine learning for extracting
  clause rules.
• Evaluation of other language pair ( eg. Japanese
  to Chinese etc)