1、The CMU Arabic-to-English Statistical MT System,Alicia Tribble, Stephan VogelLanguage Technologies Institute Carnegie Mellon University,The Data,For translation model: UN corpus: 80 million words UN Ummah Some smaller news corporaFor LM English side from bilingual corpus: Language model should have
2、seen the words generated by the translation model Additional data from Xinhua newsGeneral preprocessing and cleaning Separate punctuation mark Remove sentence pairs with large length mismatch Remove sentences which have too many non-words (numbers, special characters),The System,Alignment models: IB
3、M1 and HMM, trained in both directionsPhrase extraction From Viterbi path of HMM alignment Integrated Segmentation and AlignmentDecoder Essentially left to right over source sentence Build translation lattice with partial translations Find best path, allowing for local reordering Sentence length mod
4、el Pruning: remove low-scoring hypotheses,Some Results,Two test sets: DevTest 203 sentences, May2003 Baseline: monotone decoding RO: word reordering SL: sentence length model,Questions,Whats specific to Arabic Encoding Named Entities Syntax and Morphology Whats needed to get further improvements,Wha
5、ts Specific to Arabic,Specific to Arabic Right to left not really an issue, as this is only display Text in file is left to right Problem in UN corpus: numbers (Latin characters) sometimes in the wrong direction, eg. 1997 - 7991 Data not in vocalized form Vocalization not really studied Ambiguity ca
6、n be handled by statistical systems,Encoding and Vocalization,Encoding Different encodings: Unicode, UTF-8, CP-1256, romanized forms not too bad, definitely not as bad as Hindi;-) Needed to convert, e.g. training and testing data in different encodings Not all conversion are loss-less Used romanized
7、 form for processingConverted all data using Darwish transliteration Several characters (ya, allef, hamzda) are collapsed into two classes Conversion not completely reversibleEffect of Normalization Reduction in vocabulary: 5% Reduction of singletons: 10% Reduction of 3-gram perplexity: 5%,Named Ent
8、ities,NEs resulted in small but significant improvement in translation quality in the Chinese-English system In Chinese: unknown words are splitted into single characters which are then translated as individual words In Arabic no segmentation issues - damage less severe NEs not used so far for Arabi
9、c, but started to work on it,Language-Specific Issues for Arabic MT,Syntactic issues: Error analysis revealed two common syntactic errors Verb-Noun reordering Subject-Verb reorderingMorphology issues: Problems specific to AR morphology Based on Darwish transliteration Based on Buckwalter translitera
10、tion Poor Mans morphology,Syntax Issues: Adjective-Noun reordering,Adjectives and nouns are frequently reordered between Arabic and EnglishExample: EN: big green chair AR: chair green big Experiment: identify noun-adjective sequences in AR and reorder them in preprocessing step Problem: Often long s
11、equences, e.g. N N Adj Adj N Adj N N Result: no improvement,Syntax Issues: Subject-Noun reordering,AR: main verb at the beginning of the sentence followed by its subject EN: order prefers to have the subject precede the verbExample: EN: the President visited Egypt AR: Visited Egypt the PresidentExpe
12、riment: identify verbs at the beginning of the AR sentence and move them to a position following the first noun No full parsing Done as preprocessing on the Arabic side Result: no effect,Morphology Issues,Structural mismatch between English and Arabic Arabic has richer morphology Types Ar-En: 2.2 :
13、1 Tokens Ar-En: 0.9 : 1Tried two different tools for morphological analysis: Buckwalter analyzer http:/ competencies/content-analysis/arabic/info/buckwalter-about.html 1-1 Transliteration scheme for Arabic characters Darwish analyzer www.cs.umd.edu/Library/TRs/CS-TR-4326/CS-TR-4326.pdf Several chara
14、cters (ya, alef, hamza) are collapsed into two classes with one character representative each,Morphology with Darwish Transliteration,Addressed the compositional part of AR morphology since this contributes to the structural mismatch between AR and EN Goal was to get better word-level alignmentToolk
15、it comes with a stemmer Created modified version for separating instead of removing affixesExperiment 1: Trained on stemmed data Arabic types reduced by 60%, nearly matching number of English types But loosing discriminative power Experiment 2: Trained on affix-separated data Number of tokens increa
16、sed Mismatch in tokens much largerResult: Doing morphology monolingually can even increase structural mismatch,Morphology with Buckwalter Transliteration,Focused on DET and CONJ prefixes: AR: the, and frequently attached to nouns and adjectives EN: always separateDifferent spitting strategies: Loose
17、st: Use all prefixes and split even if remaining word is not a stem More conservative: Use only prefixes classified as DET or CONJ Most conservative: Full analysis, split only can be analyzed as a DET or CONJ prefix plus legitimate stemExperiments: train on each kind of split data Result: All set-up
18、s gave lower scores,Poor Mans Morphology,List of pre- and suffixes compiled by native speaker Only for unknown words Remove more and more pre- and suffixes Stop when stripped word is in trained lexicon Typically: 1/2 to 2/3 of the unknown words can be mapped to known words Translation not always cor
19、rect, therefore overall improvement limitedResult: this has so far been (for us) the only morphological processing which gave a small improvement,Experience with Morphology and Syntax,Initial experiments with full morphological analysis did not give an improvementMost words are seen in large corpus
20、Unknown words: 5% tokens, 10% types Simple prefix splitting reduced to halfPhrase translation captures some of the agreement informationLocal word reordering in the decoder reduces word order problemsWe still believe that morphology could give an additional improvement,Requirements for Improvements,
21、Data More specific data: We have large corpus (UN) but only small news corpora Manual dictionary could help, it helps for ChineseBetter use of existing resources Lexicon not trained on all data Treebanks not usedContinues improvement of models and decoder Recent improvements in decoder (word reordering, overlapping phrases, sentence length model) helped for Arabic Expect improvement from named entities Integrate morphology and alignment,
