syntax in statistical machine translation...statistical phrase-based translation. in proceedings of...
TRANSCRIPT
![Page 1: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/1.jpg)
Syntax in Statistical Machine Translation
Qun Liu 14 July 2015
at IIS Academia Sinica
![Page 2: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/2.jpg)
www.adaptcentre.ie Overview of Syntax in SMT
Introduc)on
Syntax-‐based Transla0on Models
Syntax-‐based Language Models
Syntax-‐based Transla0on Evalua0ons
Conclusion and Future Work
![Page 3: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/3.jpg)
www.adaptcentre.ie Google Translate - An Example
日本和美国的关系 Relationship between Japan and the United States 日本外交政策和美国亚太再平衡的关系 Japan's foreign policy and the US Asia-Pacific rebalancing of relations
![Page 4: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/4.jpg)
www.adaptcentre.ie Google Translate - An Example
日本和美国的关系 Relationship between Japan and the United States 日本外交政策和美国亚太再平衡的关系 Japan's foreign policy and the US Asia-Pacific rebalancing of relations When input sentences become longer, it is more
difficult for the Google Translate to capture their syntax structures.
![Page 5: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/5.jpg)
www.adaptcentre.ie Google Translate - An Example
Language Paris à ß English - French 91 92 English - German 77 86 English - Italian 87 89
English - Japanese 26 49 English - Chinese 17 49
Aiken, Milam, and Shilpa Balan. "An analysis of Google Translate accuracy." Translation journal 16.2 (2011): 1-3.
![Page 6: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/6.jpg)
www.adaptcentre.ie Google Translate - An Example
Language Paris à ß English - French 91 92 English - German 77 86 English - Italian 87 89
English - Japanese 26 49 English - Chinese 17 49
Aiken, Milam, and Shilpa Balan. "An analysis of Google Translate accuracy." Translation journal 16.2 (2011): 1-3.
Google Translate performs worse for language pairs with bigger difference in syntax structures.
![Page 7: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/7.jpg)
www.adaptcentre.ie Overview of Syntax in SMT
Introduc0on
Syntax-‐based Transla)on Models
Syntax-‐based Language Models
Syntax-‐based Transla0on Evalua0ons
Conclusion and Future Work
![Page 8: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/8.jpg)
www.adaptcentre.ie MT Pyramid
Vauquois’ Triangle
Analysis Generation
![Page 9: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/9.jpg)
www.adaptcentre.ie Word-based Models
IBM Model 1-5
Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The Mathematics of Statistical Machine Translation: Parameter Estimation. Computational Linguistics, 19(2):263-311.
![Page 10: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/10.jpg)
www.adaptcentre.ie Word-based Models
![Page 11: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/11.jpg)
www.adaptcentre.ie IBM Models
Source Target Probability
Bushi (布什)
Bush 0.7
President 0.2
US 0.1
yu (与)
and 0.6
with 0.4
juxing (举行)
hold 0.7
had 0.3 le
(了) hold 0.01
![Page 12: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/12.jpg)
www.adaptcentre.ie Phrase-based Models
Phrase-based Model Philipp Koehn, Franz J. Och, and Daniel Marcu. 2003. Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics Conference, pages 127-133, Edmonton, Canada, May.
Alignment Template Model Franz J. Och and Hermann Ney. 2004. The Alignment Template Approach to Statistical Machine Translation. Computational Linguistics, 30(4):417-449.
![Page 13: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/13.jpg)
www.adaptcentre.ie Phrase-based Models
![Page 14: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/14.jpg)
www.adaptcentre.ie Phrase-based Model
Source Target Probability
Bushi (布什)
Bush 0.5
president Bush 0.3
the US president 0.2
Bushi yu (布什与)
Bush and 0.8
the president and 0.2
yu Shalong (与沙⻰龙)
and Shalong 0.6
with Shalong 0.4
juxing le huiang (举行了会谈)
hold a meeting 0.7
had a meeting 0.3
![Page 15: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/15.jpg)
www.adaptcentre.ie Syntax-based Model
Dependency Syntax-‐based Model
Cons0tuent Syntax-‐based Model
Hierarchical Phrase-‐based Model
![Page 16: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/16.jpg)
www.adaptcentre.ie Hierarchical Phrase-based Model
bushi yu shalong juxing le huitan
Bush held a talk with Sharon
X 1
X 1
X 2
X 2
X 3
X 3
![Page 17: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/17.jpg)
www.adaptcentre.ie Hierarchical Phrase-based Model
Source Target Probability
juxing le huiang (举行了会谈)
hold a meeting 0.6
had a meeting 0.3
X huitang (X会谈)
X a meeting 0.8
X a talk 0.2
juxing le X (举行了X)
hold a X 0.5
had a X 0.5 Bushi yu Shalong
(布什与沙⻰龙) Bush and Sharon 0.8
Bushi X (布什X)
Bush X 0.7
X yu Y (X与Y)
X and Y 0.9
![Page 18: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/18.jpg)
www.adaptcentre.ie Hierarchical Phrase-based Model
Advantage: • Non-linguistic knowledge used
– Language Independent• High Performance
– Synchronous CFGDisadvantage:
• Limitation in long distance dependency– Use of Glue Rules for long phrases
![Page 19: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/19.jpg)
www.adaptcentre.ie Hierarchical Phrase-based Model
Advantage: • Non-linguistic knowledge used
– Language Independent• High Performance
– Synchronous CFGDisadvantage:
• Limitation in long distance dependency– Use of Glue Rules for long phrases
![Page 20: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/20.jpg)
www.adaptcentre.ie Synchronous CFG
An Introduction to Synchronous Grammars
David Chiang∗
21 June 2006
1 Introduction
Synchronous context-free grammars are a generalization of context-free grammars (CFGs) that generatepairs of related strings instead of single strings. Thus they are useful in many situations where one mightwant to specify a recursive relationship between two languages. Originally, they were developed in the late1960s for programming-language compilation [1]. In natural language processing, they have been used formachine translation [19, 20, 3] and (less commonly, perhaps) semantic interpretation.
As a preview, consider the following English sentence and its (admittedly somewhat unnatural) equiva-lent in Japanese (with English glosses):
(1) The boy stated that the student said that the teacher danced
(2) shoonen-gathe boy
gakusei-gathe student
sensei-gathe teacher
odottadanced
tothat
ittasaid
tothat
hanasitastated
One might imagine writing a finite-state transducer to perform a word-for-word translation between Englishand Japanese sentences, but not to perform the kind of reordering seen here. But a synchronous CFG can dothis.
The term synchronous CFG is recent and far from universal. They were originally known as syntax-directed transduction grammars [10] or syntax-directed translation schemata [1], the latter still probablybeing the most common name. In the NLP community they are also known as 2-multitext grammars [11],and inversion transduction grammars [19] are a special case of synchronous CFGs.
2 Definition
We give only an informal definition here. A synchronous CFG is like a CFG, but its productions have tworight-hand sides—call them the source side and the target side—that are related in a certain way. Below isan example synchronous CFG for a fragment of English and Japanese:
S→ ⟨NP 1 VP 2 ,NP 1 VP 2 ⟩(3)VP→ ⟨V 1 NP 2 ,NP 2 V 1 ⟩(4)NP→ ⟨i,watashi wa⟩(5)NP→ ⟨the box, hako wo⟩(6)V→ ⟨open, akemasu⟩(7)
∗Thanks to Philip Resnik, Anoop Sarkar, Kevin Knight, and Liang Huang for their helpful feedback.
1
The implementation of decoding algorithm is straightforward – just like a parsing procedure, either CYK or Chart algorithm works
![Page 21: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/21.jpg)
www.adaptcentre.ie Hierarchical Phrase-based Model
Advantage: • Non-linguistic knowledge used
– Language Independent• High Performance
– Synchronous CFGDisadvantage:
• Limitation in long distance dependency– Use of Glue Rules for long phrases
![Page 22: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/22.jpg)
www.adaptcentre.ie
• Using Glue Rules means sequentially concatenating all the target phrases, which lead to a back-off to phrase based model
• Two cases to use Glue Rules: Ø No hierarchical rules applicable Ø The span to be covered by the hierarchical
rule is longer than a threshold
Glue Rules
![Page 23: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/23.jpg)
www.adaptcentre.ie Glue Rules
• Using Glue Rules means sequentially concatenate all the phrases which lead to a back-off to phrase based model
• Two cases to use Glue Rules: Ø No hierarchical rules applicable Ø The span to be covered by the hierarchical
rule is longer than a threshold
Hierarchical Rules failed to capture dependency between words with a distance longer than a threshold
![Page 24: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/24.jpg)
www.adaptcentre.ie Syntax-based Model
Dependency Syntax-‐based Model
Cons0tuent Syntax-‐based Model
Hierarchical Phrase-‐based Model
![Page 25: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/25.jpg)
www.adaptcentre.ie Constituent Syntax-based Models
Forest-‐based
Tree-‐to-‐Tree
Tree-‐to-‐String String-‐to-‐Tree
![Page 26: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/26.jpg)
www.adaptcentre.ie String-to-Tree Model
● Kenji Yamada and Kevin Knight. 2001. A syntax-based statistical machine translation model. In Proceedings of ACL 2001.
● Daniel Marcu, Wei Wang, Abdessamad Echihabi, and Kevin Knight. 2006. SPMT: Statistical machine translation with syntactified target language phrases. In Proceedings of EMNLP 2006.
● Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable inference and training of context-rich syntactic translation models. In Proceedings of COLING-ACL 2006.
![Page 27: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/27.jpg)
www.adaptcentre.ie String-to-Tree Model
![Page 28: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/28.jpg)
www.adaptcentre.ie String-to-Tree Model
Source Target Probability
juxing le huiang
(举行了会谈)
VP(VPD(hold) NP(DT(a) NN(meeting))) 0.6
VP(VPD(had) NP(DP(a) NN(meeting))) 0.3
VP(VPD(had) NP(DT(a) NN(talk))) 0.1
x1 huitang (x1会谈)
VP(x1:VPD NP(DT(a) NN(meeting))) 0.8
VP(x1:VPD NP(DT(a) NN(talk))) 0.2
juxing le x1 (举行了x1)
VP(VPD(hold) NP(DT(a) x1:NN)) 0.5 VP(VPD(had) NP(DT(a) x1:NN)) 0.5
x1 yu x2 (x1与x2)
NP(x1:NNP CC(and) x2:NNP)) 0.9
![Page 29: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/29.jpg)
www.adaptcentre.ie Tree-to-String Model
● Yang Liu, Qun Liu, and Shouxun Lin. 2006. Tree-to-String Alignment Template for Statistical Machine Translation. In Proceedings of COLING/ACL 2006, pages 609-616, Sydney, Australia, July.
(Meritorious Asian NLP Paper Award) • Huang, Liang, Kevin Knight, and Aravind Joshi.
"Statistical syntax-directed translation with extended domain of locality." Proceedings of AMTA. 2006.
![Page 30: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/30.jpg)
www.adaptcentre.ie Tree-to-String Model
![Page 31: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/31.jpg)
www.adaptcentre.ie Tree-to-String Model
Source Target Probability
VPB(VS(juxing) AS(le) NPB(huiang)) (举行了会谈)
hold a meeting 0.6 have a meeting 0.3
have a talk 0.1
VPB(VS(juxing) AS(le) x1) (举行了x1)
hold a x1 0.5 have a x1 0.5
VP(PP(P(yu) x1:NPB) x2:VPB) (与 x1 x2)
x2 with x1 0.9
IP(x1:NPB VP(x2:PP x3:VPB)) x1 x3 x2 0.7
![Page 32: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/32.jpg)
www.adaptcentre.ie Tree-to-Tree Model
● Jason Eisner. 2003. Learning non-isomorphic tree mappings for machine translation. In Proc. of ACL 2003
● Min Zhang, Hongfei Jiang, Aiti Aw, Haizhou Li, Chew Lim Tan, and Sheng Li. "A tree sequence alignment-based tree-to-tree translation model." ACL-08: HLT (2008): 559.
● Yang Liu, Yajuan Lü, and Qun Liu. 2009. Improving Tree-to-Tree Translation with Packed Forests. In Proceedings of ACL/IJCNLP 2009, pages 558-566, Singapore, August.
![Page 33: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/33.jpg)
www.adaptcentre.ie Tree-to-Tree Model
![Page 34: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/34.jpg)
www.adaptcentre.ie Constituent Syntax-based Models
Advantage: • Linguistic knowledge used
Ø Long distance dependencyDisadvantage:
• Ungrammatical phrases• Syntactic Ambiguity• Computational Complexity
Ø Synchronous TSG
![Page 35: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/35.jpg)
www.adaptcentre.ie Constituent Syntax-based Models
Advantage: • Linguistic knowledge used
Ø Long distance dependencyDisadvantage:
• Ungrammatical phrases• Syntactic Ambiguity• Computational Complexity
Ø Synchronous TSG
![Page 36: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/36.jpg)
www.adaptcentre.ie Ungrammatical Phrases
• Pure tree-based models get very low performance, even lower than phrase-based models
• Various techniques are developed to incorporate ungrammatical phrases into tree-based models, which lead to an significant improvement on tree-based models He is afraid of death
R V A P N
PP
AP
VP
S
![Page 37: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/37.jpg)
www.adaptcentre.ie Constituent Syntax-based Models
Advantage: • Linguistic knowledge used
Ø Long distance dependencyDisadvantage:
• Ungrammatical phrases• Syntactic Ambiguity• Computational Complexity
Ø Synchronous TSG
![Page 38: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/38.jpg)
www.adaptcentre.ie Syntactic Ambiguity
![Page 39: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/39.jpg)
www.adaptcentre.ie 1-best ➜ n-best trees?
![Page 40: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/40.jpg)
www.adaptcentre.ie Forest-based Translation
• Mi, Haitao, Liang Huang, and Qun Liu. "Forest-Based Translation." Proceedings of ACL 2008.
• Mi, Haitao, and Liang Huang. "Forest-based translation rule extraction." Proceedings of the EMNLP 2008.
![Page 41: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/41.jpg)
www.adaptcentre.ie Packed Forest
![Page 42: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/42.jpg)
www.adaptcentre.ie N-best Trees vs. Forest
![Page 43: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/43.jpg)
www.adaptcentre.ie Constituent Syntax-based Models
Advantage: • Linguistic knowledge used
Ø Long distance dependencyDisadvantage:
• Ungrammatical phrases• Syntactic Ambiguity• Computational Complexity
Ø Synchronous TSG
![Page 44: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/44.jpg)
www.adaptcentre.ie Synchronous TSG
In a TSG derivation, we start with an elementary tree that is rooted in the start symbol, and then repeatedlychoose a leaf nonterminal symbol X and attach to it an elementary tree rooted in X. For example:
S
NP VP
V
misses
NP⇒
S
NP
John
VP
V
misses
NP(24)
⇒
S
NP
John
VP
V
misses
NP
Mary
(25)
In a synchronous TSG, the productions are pairs of elementary trees, and the leaf nonterminals are linkedjust as in synchronous CFG:
⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝
S
NP 1 VP
V
misses
NP 2
,
S
NP 2 VP
V
manque
PP
P
a
NP 1
⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠
⎛⎜⎜⎜⎜⎜⎜⎜⎜⎝NP
John,NP
Jean
⎞⎟⎟⎟⎟⎟⎟⎟⎟⎠
⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝
NP
Mary,NP
Marie
⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠
8
Synchronous CFG can be regarded as a special case of Synchronous TSG where the trees are limited to have only two layers of nodes
![Page 45: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/45.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous CFG, there is only one possible tree in the source side
![Page 46: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/46.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 47: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/47.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 48: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/48.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 49: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/49.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 50: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/50.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 51: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/51.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 52: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/52.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
Considering matching a rule starting from the root node
For Synchronous TSG, there are a large number of possible trees in the source side
![Page 53: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/53.jpg)
www.adaptcentre.ie Synchr. TSG vs Synchr. CFG
• The implementation of Synchronous TSG is much more complex than Synchronous CFG, both in space and in time
• Technologies are developed to deal with the rule indexing problem for Synchronous TSG decoder [Zhang et al., ACL-IJCNLP 2009]
• The syntax based decoder implemented in Moses does not support Synchronous TSG model with rules having more than two layers.
![Page 54: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/54.jpg)
www.adaptcentre.ie Constituent Syntax-based Models
Advantage: • Linguistic knowledge used
Ø Long distance dependencyDisadvantage:
• Ungrammatical phrases• Syntactic Ambiguity• Computational Complexity
Ø Synchronous TSGIs it possible to build a linguistically syntax-based model with the complexity of Synchronous CFG?
![Page 55: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/55.jpg)
www.adaptcentre.ie Syntax-based Model
Dependency Syntax-‐based Model
Cons0tuent Syntax-‐based Model
Hierarchical Phrase-‐based Model
![Page 56: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/56.jpg)
www.adaptcentre.ie History
Ding Y. et al. 2003, 2004 Quick C. et al. 2005 Xiong D. et al. 2007
![Page 57: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/57.jpg)
www.adaptcentre.ie Difficult of Dependency-based SMT
![Page 58: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/58.jpg)
www.adaptcentre.ie Difficult of Dependency-based SMT
…世界杯(World Cup)…在(in)…成功(Successfully) 举行(was held)
Problem: Low Coverage, Sparcity
A dependency translation rule:
![Page 59: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/59.jpg)
www.adaptcentre.ie History
Ding Y. et al. 2003, 2004 Quick C. et al. 2005 Xiong D. et al. 2007
Dependency-Treelet-based Approach
![Page 60: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/60.jpg)
www.adaptcentre.ie Dependency-Treelet-based Approach
Dependency Treelet: Any connected subgraph of a dependency tree
![Page 61: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/61.jpg)
www.adaptcentre.ie Dependency-Treelet-based Rules
![Page 62: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/62.jpg)
www.adaptcentre.ie Problem of Dep-Treelet-based Approach
• The partition of a dependency tree to a set of treelets is too flexible (more flexible than the partition of a constituent tree in a tree-to-string model)
• The reordering is difficult in target side: - These are no sequential orders between treelets- The translation of a treelet is usually non-
continuous
![Page 63: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/63.jpg)
www.adaptcentre.ie Dependency-to-String Model
• Our Solution - One layer subtree (head-dependency)- Using POS for Smoothing
Jun Xie, Haitao Mi and Qun Liu, A novel dependency-to-string model for statistical machine translation, in the Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP2011), pages 216-226, Edinburgh, Scotland, UK. July 27–31, 2011
![Page 64: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/64.jpg)
www.adaptcentre.ie Dep-to-String Rule
![Page 65: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/65.jpg)
www.adaptcentre.ie Smoothing with: Leaf nodes
![Page 66: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/66.jpg)
www.adaptcentre.ie Smoothing with: Internal nodes
![Page 67: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/67.jpg)
www.adaptcentre.ie Smoothing with: Leaf & Internal node
![Page 68: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/68.jpg)
www.adaptcentre.ie Smoothing with: Head node
![Page 69: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/69.jpg)
www.adaptcentre.ie Smoothing with: Head & Leaf nodes
![Page 70: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/70.jpg)
www.adaptcentre.ie Smoothing with: Head & Internal nodes
![Page 71: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/71.jpg)
www.adaptcentre.ie Smoothing with: All nodes
![Page 72: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/72.jpg)
www.adaptcentre.ie Experiments
![Page 73: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/73.jpg)
www.adaptcentre.ie Dependency-to-String Model
Advantage: • Linguistic knowledge used
Ø Long distance dependency• Computational Complexity
Ø Equivalent to: Synchronous CFGDisadvantage:
• Ungrammatical phrases• Syntactic Ambiguity
![Page 74: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/74.jpg)
www.adaptcentre.ie Dependency-to-String Model implemented as Synchronous CFG
Liangyou Li, Jun Xie, Andy Way, Qun Liu, Transformation and Decomposition for Efficiently Implementing and Improving Dependency-to-String Model In Moses, In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation. Pages 122-131. Doha, Qatar. 2014. • Implement Dependency-to-String in a Synchronous CFG
which is compatible with Moses chart decoder - Open Source Tools: dep2str
• Implement pseudo-forest to support partially matched head-dependency structures
![Page 75: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/75.jpg)
www.adaptcentre.ie Summary: Syntax-based Models
Syntax-‐based
Formally
ME-‐BTG Model
Hierarchical Phrase-‐Based Model
Head Transducer
Linguis0cally
cons0tuent
Tree-‐to-‐String Model
String-‐to-‐Tree Model
Tree-‐to-‐Tree Model
Dependency
Tree-‐to-‐String Model
String-‐to-‐Tree Model
Tree-‐to-‐Tree Model
![Page 76: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/76.jpg)
www.adaptcentre.ie Summary: Syntax-based Models
Syntax-‐based
Formally
ME-‐BTG Model
Hierarchical Phrase-‐Based Model
Head Transducer
Linguis0cally
cons0tuent
Tree-‐to-‐String Model
String-‐to-‐Tree Model
Tree-‐to-‐Tree Model
Dependency
Tree-‐to-‐String Model
String-‐to-‐Tree Model
Tree-‐to-‐Tree Model
Our Contribution:
![Page 77: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/77.jpg)
www.adaptcentre.ie Overview of Syntax in SMT
Introduc0on
Syntax-‐based Transla0on Models
Syntax-‐based Language Models
Syntax-‐based Transla0on Evalua0ons
Conclusion and Future Work
![Page 78: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/78.jpg)
www.adaptcentre.ie Structural Language Model
Using syntax information to capture long distance dependency in target side
![Page 79: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/79.jpg)
www.adaptcentre.ie Structural Language Model
Using syntax information to capture long distance dependency in target side
![Page 80: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/80.jpg)
www.adaptcentre.ie Existing work
• Generative Structural Language Model (Charniak, 2003) Eugene Charniak, Kevin Knight, and Kenji Yamada. 2003. Syntax-based language models for statistical machine translation. In Proceedings of MT Summit IX. Intl. Assoc. for Machine Translation.
• Idea - Estimate head n-gram probability- Using POS for smoothing
• Disadvantage - Only available when the target tree is generated- Can only be used in re-ranking rather than decoding- Generative model: features are fixed and not tunable
![Page 81: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/81.jpg)
www.adaptcentre.ie Our work
Heng Yu, Haitao Mi, Liang Huang, and Qun Liu. 2014. A Structured Language Model For Incremental Tree-to-String Translation. To be appeared in Proceedings of the 25th International Conference on Computational Linguistics (Coling2014)
• Dependency-based Language Model • Incremental: can be used in left-to-right decoding • Discriminative Model:
• Large number of used-defined features• Feature weights tunable
![Page 82: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/82.jpg)
www.adaptcentre.ie Preliminary work
• Incremental Tree-to-String Decoding Liang Huang and Haitao Mi. 2010. Efficient incremental decoding for tree-to-string translation. In Proceedings of EMNLP, pages 273–283.
• Structured perceptron with inexact search Liang Huang, Suphan Fayong, and Yang Guo. 2012. Structured perceptron with inexact search. In Proceedings of NAACL 2012, Montreal, Quebec.
![Page 83: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/83.jpg)
www.adaptcentre.ie Incremental Tree-to-String Decoding Huang 2010
![Page 84: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/84.jpg)
www.adaptcentre.ie Incremental Tree-to-String Decoding
Left-to-Right vs Top-down or Bottom-up
Huang 2010
![Page 85: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/85.jpg)
www.adaptcentre.ie Online Structural LM for SMT
![Page 86: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/86.jpg)
www.adaptcentre.ie Experiment Results
Decoding time:
8 x Baseline
Training Corpus: 1.5M sent. pairs
![Page 87: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/87.jpg)
www.adaptcentre.ie Overview of Syntax in SMT
Introduc0on
Syntax-‐based Transla0on Models
Syntax-‐based Language Models
Syntax-‐based Transla)on Evalua)ons
Conclusion and Future Work
![Page 88: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/88.jpg)
www.adaptcentre.ie Evaluation for Machine Translation
Candidate 1: It is a guide to action which ensures that the military always obeys the command of the party Candidate 2: It is to insure the troops forever hearing the activity guidebook that party direct
Reference 1: It is a guide to action that ensures that the military will forever heed party commands Reference 2: It is the guiding principle which guarantees the military forces always being under the command of the party Reference 3: It is the practical guide for the army to heed the directions of the party Question: Given the human translations as references, how to evaluation the machine translation candidates automatically?
![Page 89: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/89.jpg)
www.adaptcentre.ie Existing MT Evaluation Metrics
• Lexicalized Metrics BLEU NIST Rouge WER PER METEOR AMBER
• Syntax-based Metrics STM HWCM
• Semantic-based Metrics MEANT HMEANT
• Combinational Metrics LAYERED DISCOTK
![Page 90: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/90.jpg)
www.adaptcentre.ie Existing MT Evaluation Metrics
![Page 91: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/91.jpg)
www.adaptcentre.ie Parsing as Evaluation
• We proposed a novel MT Evaluation Metrics based on Dependency Parsing Model
• We use the reference translations as the training corpus to train a parser
• The parser are used to parse the translation candidates
• The score of the parsing model obtained by the translation candidates are regarded as its quality score.
![Page 92: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/92.jpg)
www.adaptcentre.ie WMT 2015 Metric Shared Tasks
![Page 93: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/93.jpg)
www.adaptcentre.ie Overview of Syntax in SMT
Introduc0on
Syntax-‐based Transla0on Models
Syntax-‐based Language Models
Syntax-‐based Transla0on Evalua0ons
Conclusion and Future Work
![Page 94: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/94.jpg)
www.adaptcentre.ie Conclusion
• Syntax-based Translation Models • Constituent Tree-to-String Model• Forest-based Translation Approach• Dependency-based Model
• Syntax-based Language Model • Online Discriminative Structural LM for SMT
• Syntax-based Translation Evaluation Metrics • Dependency Parsing as Evaluation for SMT
![Page 95: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/95.jpg)
www.adaptcentre.ie Future Work
• Graph-based Translation Model - Sequence-based è Tree-based è Graph-based- A natural framework to incorporate various linguistic
knowledge (1) n-gram (2) morphology (3) syntax (4) semantic
• Dependency Parsing as Evaluation for SMT - Extension to a discriminative model- Used as a combination framework
![Page 96: Syntax in Statistical Machine Translation...Statistical Phrase-Based Translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics](https://reader033.vdocuments.site/reader033/viewer/2022050405/5f8300a8dc3f0145be040fe4/html5/thumbnails/96.jpg)
Q&A !Qun Liu, Professor, Dr.,PI
ADAPT Centre Dublin City University
Institute of Computing Technology Chinese Academy of Sciences
Email: [email protected]