Statistical PhraseBased Translation Authors: Koehn, Och, Marcu

1
Statistical Phrase-Based TranslationAuthors
Koehn, Och, Marcu

• Presented by Albert Bertram
• Titles, charts, graphs, figures and tables were
  extracted from the paper.
• Critique and snarky remarks, however, are
  original.

2
Motivation

• We have a new way of learning phrase
  translations, but
• What is the best method to extract phrase
  translation pairs?

3
What to do, what to do

• Compose a framework for consistent comparison
• Implement each algorithm
• Compare the results

4
Evaluation Framework

• Phrases
• Models involved
• Language model
• Statistical model for translation
• Distortion model
• Decoder

5
Evaluation Framework Phrases

• We all know what phrases are right?
• NP, VP, wait, what? Oh.
• Here, theyre generic spanning and
  non-overlapping subsequences of words.
• Are these guys really linguists?

6
Evaluation Framework Models

• Language Model
• Trigram usually p(enen-1,en-2)
• Translation Model
• Argmaxe p(ef) argmaxe p(fe)p(e)
• ebest argmaxe p(fe)pLM(e)?length(e)
• p(fe) is decomposed into

7
Evaluation Framework Models

• Distortion Model
• d(ai bi-1)
• ai start position of the foreign phrase
  translated into the ith English phrase
• bi-1 end position of the foreign phrase
  translated into the (i-1)th English phrase
• Learned from the joint probability model Ping
  told us about

8
Evaluation Framework Decoder

• Left-to-right incremental
• Stack-based beam search
• Estimates future costs
• Same decoder used in all experiments

9
Baseline Experiments

• Word-based alignment
• Syntactic phrases
• Phrase alignments

10
Baseline Word-based Alignment

• Learn the phrases from word alignments

11
Baseline Syntactic Phrases

• Learn only syntactically correct phrases
• Start with the word based alignment
• Prune out the phrase pairs which arent subtrees
  in the parsed sentences for either language.

12
Baseline Phrase Alignment

• Marcu and Wong, 2002
• Yes, this is the paper Ping just presented.

13
Experiment Background

• Europarl and BLEU
• Training corpus of 10, 20, 40, 80,160 and 320
  kilo-sentence pairs

14
Baseline Results
Notice the bottom row there? Comparing these
models is like taking a 5-year old to a chess
tournament.
15
Baseline Results
16
More Experiments

• Weighting Syntactic Phrases
• Maximum Phrase Length
• Lexical Weighting
• Phrase Extraction Heuristic
• Simpler Underlying Word-Base Models
• Other Languages

17
Experiments and Results

• Weighting Syntactic Phrases
• Double the count on syntactic phrases
• Is that sufficient?
• Insufficient post-analysis on this one
• The same BLEU score
• Were the translations in better syntax?
• Did the translations at least use more syntactic
  phrases?

18
Experiments and Results

• Maximum Phrase Length

19
Experiments and Results

• Lexical Weighting
• Lexical probability distribution
• Lexical Weight

20
Experiments and Results

• Lexical Weighting
• Example

21
Experiments and Results

• Lexical Weighting
• With multiple alignments
• Extended to fit this model

22
Experiments and Results

• Lexical Weighting
• Improvement .01 BLEU

23
Experiments and Results

• Phrase Extraction Heuristic
• Align Bidirectionally
• Note this gives two different word alignment sets
• Start with the intersection of the two sets
• Add possible alignments
• Only if theyre in the union of the sets
• Only if they connect at least one previously
  unaligned word

24
Experiments and Results

• Phrase Extraction Heuristic
• Algorithm
• Start with the first English word
• Expand only directly adjacent alignment points
• Move to the next English word, repeat.
• Finally add non-adjacent alignment points which
  meet the heuristic criteria.

25
Experiments and Results
26
Experiments and Results

• Simpler Underlying Word-Base Models
• IBM models 1-4

27
Experiments and Results

• Other Languages