Modality-Preserving Phrase-Based Statistical Machine Translation

Masamichi Ideue, Masao Utiyama, Eiichiro Sumita and Kazuhide Yamamoto. Modality-Preserving Phrase-Based Statistical Machine Translation. Proceedings of the International Conference on Asian Language Processing (IALP 2012), pp.129-132 (2012.11)

1. .PEBMJUZ1SFTFSWJOH 1ISBTF#BTFE 4UBUJTUJDBM
   .BDIJOF
   5SBOTMBUJPO Masamichi Ideue, Masao Utiyama, Eiichiro Sumita and Kazuhide Yamamoto (Nagaoka University of Technology and NICT)
2. Purpose of our study Japanese to English translation preserving negation and question modality by Phrase- based SMT. Input ࢲ͸ΓΜ͕͝޷͖Ͱ͸͋Γ·ͤΜɻ Translation I don’t like apples. The MT users would not be able to detect a modality error.  MT Translation I like apples.
3. Related Studies • Class-Dependent Modeling for Dialog Translation [Finch et al., 2009] • Discriminative Reranking for SMT using Various Global Features [Goh et al., 2010] Our study focused on characteristic modality words in negations and questions. Neither of the studies discussed what expressions influence modalities. 
4. Proposed Method Add feature functions considered characteristic words of negation and question. 
5. Added feature functions The number of phrase pairs including characteristic words of question (negation) in Japanese phrase and English phrase. )ZQPUIFTJT
   e
   8IFSF
   
   JT
   
   UIF
   
   QVSTF
   
   *OQVU
   f
   ࡒ෍
   
   ͸
   
   Ͳ͜
   
   ʹ
   
   ͋Γ
   ·͢
   
   ͔
   
   ʁ  
6. Characteristic Words Extraction • Manual extraction • Automatic extraction • Using LLR(Log-likelihood ratio) score Extract characteristic words from the parallel corpus in travel domain. 
7. Manual Extraction &OHMJTI /FHBUJPO 2VFTUJPO OPU
   `U EPO %PO IBWFO JTO /P XPO XBTO EPFTO EJEO DBOOPU IBEO 8IZ 8JMM 8IBU $PVME *T )PX %PFT $BO %P "SF 8IJDI 8IFO 8IFSF )BWF %PFT %JE 8BT .BZ 
8. Manual Extraction +BQBOFTF /FHBUJPO 2VFTUJPO ͳ͍ OBJ ·ͤΜ NBTFO ͔
   ɻ LB • The characteristic words that clearly express the modalities are few. • Whether a word expresses modality or not, there is tendency to depends on the domain. 
9. Automatic Extraction • Automatic extraction is based on LLR. • LLR is convenient for extracting characteristic words in travel domain (Chujo et al., 2006).   8JMM  $PVME  )PX  $BO   Extract top N words from the ranking by LLR score as the characteristic words. 0SEFS
   CZ
   --3
   TDPSF
   2VFTUJPO 
10. Calculation of LLR *O
    DBTF
    PG
    OFHBUJPO *G
    B
    XPSE
    UFOET
    UP
    PDDVS
    JO
    OFHBUJPO
    POMZ
    UIF
    --3
    TDPSF
    CFDPNFT
    IJHI
    /FHBUJPO "⒏SNBUJPO X B C B C X D E D E B D C E O B C D E
    
    PDDVSSFODF
    GSFRVFODZ
    JO
    FBDI
    DPOEJUJPO 
11. Sentence type classification 5P
    CVJME
    UIF
    DPOUJOHFODZ
    UBCMF
    XF
    EJWJEFE
    TFOUFODFT
    JO
    UIF
    QBSBMMFM
    DPSQVT
    XJUI
    NBOVBMMZ
    FYUSBDUFE
    &OHMJTI
    DIBSBDUFSJTUJD
    XPSET &OHMJTI +BQBOFTF 5ZQF )F
    JT
    OPU
    BO
    BSUJTU ൴͸ܳज़ՈͰ͸ͳ͍ɻ OFHBUJPO *
    MJLF
    BQQMFT ࢲ͸ΓΜ͕͝޷͖Ͱ͢ɻ B⒏SNBUJPO "SF
    ZPV
    B
    EPDUPS ͋ͳͨ͸ҩऀͰ͔͢ɻ RVFTUJPO 
12. Extracted Words by LLR &OHMJTI /FHBUJPO 2VFTUJPO EP BOZ UIFSF IBWF UIJT EPO MPOH JU JTO EJE ZPVS NVDI IPX UJNF DBO ZFU BOZ CVU LOPX XPSSZ * BOZUIJOH JU TP BGSBJE VOEFSTUBOE XIBU FOPVHI 
13. Extracted Words by LLR +BQBOFTF /FHBUJPO 2VFTUJPO ͔ Ͳ͜ Կ Ͳ͏ ͍͘Β ͸ ͍͚ͨͩ Ͳͷ Կ࣌ ͋Γ Ͱ͠ΐ ΋Β͑ ͍͔͕ ͲΜͳ ·ͤ ͳ͍ Μ ͸ ͳ͔ͬ ͋·Γ ·ͩ ͋Γ Ͱ͖ ͡Ό ͍͍͑ ͦΜͳʹ ͦΜͳ ͨ͘ 
14. Experiments 4.5
    5PPMLJU .PTFT 5VOJOH .JOJNVN
    &SSPS
    3BUF
    5SBJOJOH 1BSBMMFM
    DPSQVT #BTJD
    5SBWFM
    &YQSFTTJPO
    $PSQVT
    #5&$
     
    QBJST 5FTU
    TFU  
    TFOUFODFT JODMVEFE
    
    TFOUFODFT
    GPS
    OFHBUJPO
    RVFTUJPO
    BOE
    B⒏SNBUJPO %FWFMPQNFOU
    TFU  
    TFOUFODFT
    JO
    UIF
    TBNF
    XBZ
    BT
    UFTU
    TFU 
15. Experiments • From preliminary experimental evaluation with BLEU, the N is decided as 30 (LLR30). • Baseline method is no additional features. 
16. Manual Evaluation • To verify effectiveness of translation quality when add the proposed features. • To verify accuracy of each modality. 8F
    SBOEPNMZ
    FYUSBDUFE
    
    QBJST
    UP
    UFTU
    UIF
    NFUIPET
    GPS
    FBDI
    NPEBMJUZ
    UPUBM
    
    QBJST  
17. Translation Quality (PPE 4 " # 4 " # $ % #BTFMJOF /P
    BEEJUJPOBM
    GFBUVSFT       .BOVBMMZ
    &YUSBDUJPO       --3       "MM
    UIF
    NFUIPET
    IBWF
    UIF
    TBNF
    USBOTMBUJPO
    RVBMJUZ
    JG
    4
    "
    BOE
    #
    BSF
    BTTVNFE
    HPPE
    USBOTMBUJPO OVNCFS
    PG
    TFOUFODFT 
18. Accuracy of each modality "⒎ /FH 2VF #BTFMJOF    .BOVBMMZ
    &YUSBDUJPO    --3    1FSDFOUBHF
    PG
    UIF
    PVUQVUT
    QSFTFSWFE
    UIF
    NPEBMJUZ
    PG
    UIF
    JOQVU w1SPQPTFE
    NFUIPET
    JOEJDBUFE
    B
    NBSLFE
    JNQSPWFNFOU
    JO
    OFHBUJPO
    NPEBMJUZ w5IF
    BDDVSBDZ
    PG
    --3
    XBT
    CFUUFS
    UIBO
    UIF
    BDDVSBDZ
    
    PG
    UIF
    CBTFMJOF
    JO
    BMM
    NPEBMJUJFT 
19. Translation Example 1SPQPTFE
    NFUIPE
    .BOVBMMZ
    &YUSBDUJPO  8IJDI
    POF
    TIBMM
    XF
    HP
    UP
    UIF
    DJSDVT
    BOE
    [PP
    0 *OQVU
    2VFTUJPO  αʔΧεͱಈ෺ԂɺͲͬͪʹߦ͜͏͔ɻ #BTFMJOF -FU`T
    HP
    UP
    UIF
    DJSDVT
    BOE
    UIF
    [PP
    9 
20. Translation Example 1SPQPTFE
    NFUIPE
    .BOVBMMZ
    &YUSBDUJPO  *
    EPO`U
    NJOE
    JG
    ZPV
    DBODFM
    JU
    9 *OQVU
    2VFTUJPO  Ωϟϯηϧͯ͠΋͔·͍·ͤΜ͔ɻ #BTFMJOF .BZ
    *
    DBODFM
    0 NBTFO OFHBUJPO NBTFO
    LB RVFTUJPO 8F
    IBWF
    UP
    USFBU
    XPSE
    DPNCJOBUJPOT 
21. Conclusion • We proposed additional feature considering characteristic words for modality-preserving PBSMT. • Produced more translations preserved the modality of the input sentence than baseline without decrease of translation quality. • Automatic extraction performed the same as or better than manual extraction. 
22. LLR
23. LLR /FHBUJPO "⒏SNBUJPO X B C B C X D E D E B D C E O
24. Translation Example 1SPQPTFE
    NFUIPE
    .BOVBMMZ
    &YUSBDUJPO  1MFBTF
    HP
    FBTZ
    0 *OQVU
    "⒏SNBUJPO  ΍͘͞͠ଧ͍ͬͯͩ͘͞Ͷɻ 1SPQPTFE
    NUIPE
    &OHMJTI
    TJEF
    POMZ  1MFBTF
    HP
    FBTZ
    JTO`U
    JU
    9
25. Calculation of LLR • Pr(D|H_indep) is the probability under the null hypothesis that the occurrences of a word w in the negative and affirmative sentences are independent of one another. • Pr(D|H_dep) is the case in which the occurrences are dependent. *O
    DBTF
    PG
    OFHBUJPO *G
    B
    XPSE
    UFOET
    UP
    PDDVS
    JO
    OFHBUJPO
    POMZ
    UIF
    --3
    TDPSF
    CFDPNFT
    IJHI
    
26. Calculation of LLR /FHBUJPO "⒏SNBUJPO X B C B C X D E D E B D C E O B C D E
    
    PDDVSSFODF
    GSFRVFODZ
    JO
    FBDI
    DPOEJUJPO
27. Related Studies • Class-Dependent Modeling for Dialog Translation [Finch et al., 2009] • 2 models are trained for question sentence and other sentence. • Discriminative Reranking for SMT using Various Global Features [Goh et al., 2010] • Probabilities of sentence types such as negations and questions are used. Our study focused on characteristic modality words in negations and questions. Neither of the studies discussed what expressions influence modalities.