Learning Translation Templates and Rules from Bilingual Knowledge Bank in ExampleBased Machine Trans

1
Learning Translation Templates and Rules
fromBilingual Knowledge Bank in Example-Based
Machine Translation

• Prepared by Ye Hong Hoe

2
Introduction

• Example-Based Machine Translation (EBMT)
• translation examples are used to translate
• involve three steps
• decompose an input sentence into phrases
• translate the phrases into target language
  phrases
• combine the target phrases into one sentence

3
Introduction (cont.)

• Bilingual Knowledge Bank (BKB)
• contains a set of translation examples
• syntactically structured
• translation units encoded

4
Introduction (cont.)

• Translation Template
• generalized translation example
• for example,
• English a XAdj difference
• Malay perbezaan XAdj

5
Introduction (cont.)

• Translation Rule
• contains only POSs or syntactic categories
• for example,
• English Det Adj N
• Malay N Adj

6
Motivations

• Broader Coverage
• large parallel corpus is difficult to get
• by using templates and rules, same coverage can
  be achieved with smaller corpus
• Grammatical Accuracy
• by using templates and rules, syntactic
  structures can be kept

7
Objectives

• To learn translation templates and rules from BKB
• To develop an English-Malay EBMT system

8
Outline

• Background
• Methodology
• Preliminary Result
• Discussion
• Conclusion

9
Background

• Structured String-Tree Correspondence (SSTC)
• a general structure, where a string is associated
  with its representation tree structure

English
knelt v (1_2/0_5)
knelt v (1_2/0_5)
he pron (0_1/0_1)
on prep (2_3/2_5)
floor n (4_5/3_5)
the det (3_4/3_4)
0he1knelt2on3the4floor5
Figure 1. An example SSTC (simplified)
10
Background (cont.)

• Synchronous SSTC (S-SSTC)
• consists of a pair of SSTCs with synchronous
  correspondences between them

English
Malay
knelt v (1_2/0_5)
knelt v (1_2/0_5)
berlutut v (1_2/0_6)
he pron (0_1/0_1)
on prep (2_3/2_5)
dia pron (0_1/0_1)
di atasprep (2_4/2_6)
floor n (4_5/3_5)
lantai n (4_5/4_6)
the det (3_4/3_4)
itu det (5_6/5_6)
0he1knelt2on3the4floor5
0dia1berlutut2di3atas4lantai5itu6
Figure 2. An example S-SSTC (simplified)
11
Background (cont.)

• EBMT Based on S-SSTC (Al-Adhaileh Tang, 1999)
• Process
• based on word index, select relevant examples
• based on correspondence, build a list of sub
  S-SSTCs
• based on a best example, combine the sub S-SSTCs

12
Background (cont.)

• Weaknesses
• only index at word level
• example result knelt on ? berlutut kepada
• cannot handle unknown words
• example result knelt on the floor ? berlutut
  kepada lantai
• only use template at top level
• example result

13
Methodology

• System Overview
• Indexing of Translation Examples
• Learning Translation Templates and Rules
• Translation Using Examples, Templates, and Rules

14
System Overview
15
Indexing

• Indexing at structural level
• Indexing at word level

16
Learning Templates

• Types of Templates
• Type 1 simple template
• Type 2 template with extended context
• Type 3 template with partial tree correspondence

17
Learning Template Type 1

• Simple Template
• contains only one level of child node(s) in its
  representation trees

18
Learning Template Type 2

• Template with Extended Context
• context of the template is extended downwards
  until content word

knelt v
berlutut v
X pron
on prep
X pron
di atas prep
Y n
Y n
X knelt on Y
X berlutut di atas Y
Figure 7. An example of template type 2
19
Learning Template Type 3

• Template with Partial Tree Correspondence
• partial tree correspondence exists between
  representation trees

20
Learning Rules

• Rules
• extracted from template type 1

Y v
Y v
X pron
Z prep
X pron
Z prep
Pron V Prep
Pron V Prep
X prep
X n
X prep
X n
Y n
Y det
Y n
Y det
Prep N
Det N
Prep N
N Det
Figure 9. Examples of rules
21
Translation

• Example Matching
• Recombination
• Using Templates
• Using Rules

22
Example Matching

• Match tagged source sentence against translation
  examples in indexed BKB
• for example,

23
Recombination

• Combine source / target chunks using rules
• for example,

24
Recombination (cont.)

• Combine source/target chunks using templates
• for example,

25
Preliminary Result
Figure 13. Comparison between our approach and
previous approach for 20 input
sentences
26
Discussion

• Al-Adhaileh and Tangs approach has better result
  because
• our approach found wrong target words for source
  function words
• e.g. is ada, was - sedang
• our approach only handled exact match
• However,
• our approach produced better sentence structures
• our approach looked up unknown word in bilingual
  lexicon

27
Conclusion

• Translation templates and rules
• has extended coverage of our system
• can be used to recombine source and target chunks
  with proper syntactic structure
• However, there is a need
• to have better handling of function words
• to explore possibility of doing fuzzy match

28
Thank YouTerima Kasih