Statistical Phrase-Based Translation Authors: Koehn, Och, Marcu Presented by Albert Bertram Titles, charts, graphs, figures and tables were extracted from.

Statistical Phrase-Based TranslationAuthors: Koehn, Och, Marcu

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

Motivation

We have a new way of learning phrase translations, but…

“What is the best method to extract phrase translation pairs?”

What to do, what to do

Compose a framework for consistent comparison

Implement each algorithmCompare the results

Evaluation Framework

PhrasesModels involved

Language modelStatistical model for translationDistortion model

Decoder

Evaluation Framework: Phrases

We all know what phrases are right?NP, VP, wait, what? Oh.

Here, they’re generic spanning and non-overlapping subsequences of words.Are these guys really linguists?

Evaluation Framework: Models

Language Model Trigram usually p(en|en-1,en-2)

Translation Model Argmaxe p(e|f) = argmaxe p(f|e)p(e)

ebest = argmaxe p(f|e)pLM(e)ωlength(e)

p(f|e) is decomposed into

Evaluation Framework: Models

Distortion Modeld(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

Evaluation Framework: Decoder

Left-to-right incrementalStack-based beam searchEstimates future costsSame decoder used in all experiments

Baseline Experiments

Word-based alignmentSyntactic phrasesPhrase alignments

Baseline: Word-based Alignment

Learn the phrases from word alignments

Baseline: Syntactic Phrases

Learn only syntactically correct phrasesStart with the word based alignment Prune out the phrase pairs which aren’t

subtrees in the parsed sentences for either language.

Baseline: Phrase Alignment

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

presented.

Experiment Background

Europarl and BLEUTraining corpus of 10, 20, 40, 80,160

and 320 kilo-sentence pairs

Baseline Results

Notice the bottom row there? Comparing these models is like taking a 5-year old to a chess tournament.

Baseline Results

More Experiments

Weighting Syntactic PhrasesMaximum Phrase LengthLexical WeightingPhrase Extraction HeuristicSimpler Underlying Word-Base ModelsOther Languages

Experiments and Results

Weighting Syntactic PhrasesDouble the count on syntactic phrases

Is that sufficient? Insufficient post-analysis on this oneThe same BLEU score

Were the translations in better syntax?Did the translations at least use more syntactic

phrases?

Experiments and Results

Maximum Phrase Length

Experiments and Results

Lexical WeightingLexical probability distribution

Lexical Weight

Experiments and Results

Lexical WeightingExample

Experiments and Results

Lexical WeightingWith multiple alignments

Extended to fit this model

Experiments and Results

Lexical Weighting Improvement: .01 BLEU

Experiments and Results

Phrase Extraction HeuristicAlign Bidirectionally

Note this gives two different word alignment setsStart with the intersection of the two setsAdd possible alignments

Only if they’re in the union of the setsOnly if they connect at least one previously

unaligned word

Experiments and Results

Phrase Extraction HeuristicAlgorithm

Start with the first English wordExpand only directly adjacent alignment pointsMove to the next English word, repeat.Finally add non-adjacent alignment points which

meet the heuristic criteria.

Experiments and Results

Experiments and Results

Simpler Underlying Word-Base Models IBM models 1-4

Experiments and Results

Other Languages