MT with Limited Resources: Approaches and Results

1
MT with Limited ResourcesApproaches and Results

• Ralf Brown, Stephan Vogel, Alon Lavie, Lori
  Levin, Jaime Carbonell
• Students Christian Monson, Erik Peterson,
  Kathrin Probst, Ashish Venugopal, Ying Zhang
• Carnegie Mellon University

2
RADD CMU's "incubator" for MT technologies

• Multiple Techniques
• Statistical
• Example-Based
• Transfer-Rule
• Common Pre-Processing
• Segmentation
• Conversion of numbers to Arabic numerals
• Translation of month names to English month names
• Multi-Engine combinations

3
Statistical MT

• The major improvement for June evaluation was
  phrase-to-phrase alignments.
• Performance (NIST score)ME-Compatible-SMT
  5.7354Full-SMT 6.1361

4
Example-Based MT

• Given an indexed training corpus,
• find phrases in the corpus which occur in the
  input to be translated
• retrieve the sentence pairs containing matches,
  and
• perform a word-level alignment to determine
  translations.
• Our "standard" EBMT system is actually a
  multi-enginecombination of phrasal EBMT, LDC
  lexicon, and statistical dictionary extracted
  from the training text.

5
Example-Based MT (2)

• Inexact matching A phrase can match even if one
  of the words inside the phrase differs, provided
  that the dictionary can provide a more-or-less
  unambiguous translation for the unmatched word.
• "More-or-less Unambiguous" means that the
  translationwith the second-highest frequency has
  frequency less thanTHRESHOLDhighest-frequency
• we experimentally determined the best threshold
  to be 0.55.

6
Transfer Rule MT

• Manually-developed transfer rules for translation
  with our newly developed Transfer Engine.
• (71 hours development time)
• and a transfer lexicon automatically derived from
  theLDC 10k-word lexicon.
• A language model selected from among ambiguous
  translations.
• Performance (NIST score)XFERLM 4.8404

7
Multi-Engine MT

• Hypothesis by combining multiple
  translationmethods, we can mitigate weaknesses
  and enhancestrengths of individual methods.
• Each engine generates whatever partial
  translationsit can and assigns an approximate
  quality score.
• The partial translations are then combined into
  alattice and a trigram model of the output
  language(plus other scoring heuristics) is used
  to selectthe best path through the lattice.

8
Multi-Engine MT Results

• Most combinations outperform the individual
  engines.
• We submitted two combinations and the engines
  they combinedfor official scores, with the
  result shown below.PhrEBMT 3.9668 PhrEBMT 3.96
  68SMT 5.7354 XFER
  4.8404Combo 5.9524 Combo 5.2170
• Additionally, we see the effect of
  combininglexica with phrasal EBMTPhrEBMT
  3.9668PhrEBMTlex 5.2883

9
CMU Small Data Results

• Official results were submitted with segmentor
  trained on full LDC word list (same as large
  data track)
• We retrained our segmentor with only 10K small
  dict and the words from the 100K Chinese treebank
  and re-evaluated our latest systems.
• Results with new (small) segmentation are
  reported in parentheses

10
CMU Small Data Results SMT

• Results with full segmentor and with re-trained
  small segmentor for different versions of our
  SMT system

11
Learning Transfer-Rules for Languages with
Limited Resources

• Rationale
• Large bilingual corpora not available
• Bilingual native informant(s) can translate and
  align a small pre-designed elicitation corpus,
  using elicitation tool
• Elicitation corpus designed to be typologically
  comprehensive and compositional
• Transfer-rule engine and new learning approach
  support acquisition of generalized transfer-rules
  from the data

12
AVENUE Transfer
13
Sample Transfer Rule

• Rules contain necessary information for analysis,
  transfer and generation
• Unification equations used to build source,
  target feature structures
• Transfer Chinese questions formed by appending
  particle MA to English
• S,2 Rule ID
• SS NP VP MA -gt AUX NP VP Source
  Target production rules
• (
• (x1y2) Source NP
  aligns with target NP
• (x2y3) Source VP
  aligns with target VP
• ((x0 subj) x1) Build the source
  feature structure
• ((x0 subj case) nom)
• ((x0 act) quest)
• (x0 x2)
• ((y1 form) do) Set inserted constituent
  AUXs base form to do
• ((y3 vform) c inf) Constrain verb to
  infinitive form
• ((y1 agr) (y2 agr)) Enforce agreement
  between do and subject
• )

14
Transfer Overview

• The AVENUE translation engine developed
  internally and follows a three-step transfer
  approach
• Analysis
• Transfer
• Generation
• The engine can be run with manually developed
  transfer-rules as a stand-alone system or operate
  as part our larger rule-learning system.

15
RADD Transfer Development

• Total Chinese-specific rule and lexicon
  development time 71 hours
• Small and Large Tracks used same transfer rules
  but different sized lexicons (10K vs. 50K)
• Rule development by a bilingual speaker with
  linguistic background, based upon manual
  evaluation of training data and personal
  grammatical knowledge.
• Development concentrated on translating noun
  phrases and structures where Chinese and English
  word order differed

16
Analysis

• Analysis uses a uses a unification-based chart
  parser to find the input sentences grammatical
  structure.
• Different possible analyses and transfer paths
  are all efficiently packed together in a packed
  forest for later usage.

17
Transfer

• Transfer rule manipulate the parse tree(s)
  created during analysis.
• Constituents (such as noun, verb phrases) can be
  reordered, inserted, or deleted.
• Words are translated using a transfer lexicon
• For sentences without a complete parse, transfer
  occurs on the longest sub-parses found during
  analysis.

18
Generation

• During generation, the engine checks that the
  target language tree from transfer satisfies
  target language constraints (e.g. subject-verb
  agreement in English)
• Finally, the target sentence is read from the
  leaves of the target tree and returned.

19
Rule Learning - Overview

• Goal Acquisition of Syntactic Transfer Rules
• 1) Flat Seed Generation produce rules from
  word-aligned sentence pairs, abstracted only to
  POS level no syntactic structure
• 2) Add compositional structure to Seed Rule by
  exploiting previously learned rules
• 3) Seeded Version Space Learning group seed rules
  by constituent sequences and alignments, seed
  rules form s-boundary of VS generalize with
  validation

20
Flat Seed Generation

• Create a seed rule that is specific to the
  sentence pair, but abstracted to the pos level.
  Use SL information (e.g. parses), and any TL
  information. E.g.
• The highly qualified applicant visited the
  company.
• Der äußerst qualifizierte Bewerber besuchte die
  Firma.
• ((1,1),(2,2),(3,3),(4,4),(5,5),(6,6))

SS det adv adj n v det n? det adv adj n v
det n (alignments (x1y1) (x2y2) (x3y3)
(x4y4) (x5y5) (x6y6) (x7y7) constraints
((x1 def) ) ((x4 agr) 3-sing) ((x5 tense)
past) . ((y1 def) ) ((y3 case) nom)
((y4 agr) 3sg) )
21
Compositionality

• If there is a previously learned rule that can
  account for part of the sentence, adjust seed
  rule to reflect this compositional element.
• Adjust constituent sequences, alignments, and
  constraints add context constraints (from
  possible translations), remove unnecessary ones

SS det adv adj n v det n? det adv adj n v
det n (alignments (x1y1) (x2y2) (x3y3)
(x4y4) (x5y5) (x6y6) (x7y7) constraints
((x1 def) ) ((x4 agr) 3-sing) ((x5 tense)
past) . ((y1 def) ) ((y4 agr) 3sg)
)
NPNP det adv adj n det adv adj
n ((x1y1) ((y4 agr) (x4 agr) .)
SS NP v det n? NP v det n (alignments (x1
y1) (x2y2) (x3y3) (x4y4) (x5y5)
(x6y6) (x7y7) constraints ((x5 tense)
past) . ((y1 def) ) ((y1 case) nom)
((y1 agr) 3sg) )
22
Seeded Version Space Learning

• NP v det n NP VP
• Group seed rules into version spaces as above.
• Make use of partial order of rules in version
  space. Partial order is defined
• via the f-structures satisfying the constraints.
• Generalize in the space by repeated merging of
  rules
• Deletion of constraint
• Moving value constraints to agreement
  constraints, e.g.
• ((x1 num) pl), ((x3 num) pl) ?
• ((x1 num) (x3 num)
• 4. Check translation power of generalized rules
  against sentence pairs

23
Future Work

• Baseline evaluation
• Adjust generalization step size
• Revisit generalization operators
• Introduce specialization operators to retract
  from overgeneralizations (including seed rules)
• Learn from an unstructured bilingual corpus
• Evaluate merges to pick the optimal one at any
  step based on cross-validation, number of
  sentences it can translate