Rapid Prototyping of Machine Translation Systems A Tale of Two Case Studies Srinivas Bangalore Giuseppe Riccardi AT&T Labs-Research Joint work with German.

Slides:



Advertisements
Similar presentations
The Application of Machine Translation in CADAL Huang Chen, Chen Haiying Zhejiang University Libraries, Hangzhou, China
Advertisements

Statistical Machine Translation Part II: Word Alignments and EM Alexander Fraser Institute for Natural Language Processing University of Stuttgart
Statistical Machine Translation Part II: Word Alignments and EM Alexander Fraser ICL, U. Heidelberg CIS, LMU München Statistical Machine Translation.
Statistical Machine Translation Part II – Word Alignments and EM Alex Fraser Institute for Natural Language Processing University of Stuttgart
Proceedings of the Conference on Intelligent Text Processing and Computational Linguistics (CICLing-2007) Learning for Semantic Parsing Advisor: Hsin-His.
MT Evaluation: Human Measures and Assessment Methods : Machine Translation Alon Lavie February 23, 2011.
A Syntactic Translation Memory Vincent Vandeghinste Centre for Computational Linguistics K.U.Leuven
Improving Machine Translation Quality via Hybrid Systems and Refined Evaluation Methods Andreas Eisele DFKI GmbH and Saarland University Helsinki, November.
For Friday No reading Homework –Chapter 23, exercises 1, 13, 14, 19 –Not as bad as it sounds –Do them IN ORDER – do not read ahead here.
Languages & The Media, 4 Nov 2004, Berlin 1 Multimodal multilingual information processing for automatic subtitle generation: Resources, Methods and System.
Inducing Information Extraction Systems for New Languages via Cross-Language Projection Ellen Riloff University of Utah Charles Schafer, David Yarowksy.
1 Noun Homograph Disambiguation Using Local Context in Large Text Corpora Marti A. Hearst Presented by: Heng Ji Mar. 29, 2004.
ASR Evaluation Julia Hirschberg CS Outline Intrinsic Methods –Transcription Accuracy Word Error Rate Automatic methods, toolkits Limitations –Concept.
The Use of Speech in Speech-to-Speech Translation Andrew Rosenberg 8/31/06 Weekly Speech Lab Talk.
Machine Translation Prof. Alexandros Potamianos Dept. of Electrical & Computer Engineering Technical University of Crete, Greece May 2003.
Course Summary LING 575 Fei Xia 03/06/07. Outline Introduction to MT: 1 Major approaches –SMT: 3 –Transfer-based MT: 2 –Hybrid systems: 2 Other topics.
تمرين شماره 1 درس NLP سيلابس درس NLP در دانشگاه هاي ديگر ___________________________ راحله مکي استاد درس: دکتر عبدالله زاده پاييز 85.
Semi-Automatic Learning of Transfer Rules for Machine Translation of Low-Density Languages Katharina Probst April 5, 2002.
Comments on Guillaume Pitel: “Using bilingual LSA for FrameNet annotation of French text from generic resources” Gerd Fliedner Computational Linguistics.
1 The Web as a Parallel Corpus  Parallel corpora are useful  Training data for statistical MT  Lexical correspondences for cross-lingual IR  Early.
1 Statistical NLP: Lecture 13 Statistical Alignment and Machine Translation.
A Pattern Matching Method for Finding Noun and Proper Noun Translations from Noisy Parallel Corpora Benjamin Arai Computer Science and Engineering Department.
DIVINES – Speech Rec. and Intrinsic Variation W.S.May 20, 2006 Richard Rose DIVINES SRIV Workshop The Influence of Word Detection Variability on IR Performance.
Natural Language Processing Expectation Maximization.
Lecture 1, 7/21/2005Natural Language Processing1 CS60057 Speech &Natural Language Processing Autumn 2005 Lecture 1 21 July 2005.
Stochastic Transductions for Machine Translation*
Natural Language Processing Lab Northeastern University, China Feiliang Ren EBMT Based on Finite Automata State Transfer Generation Feiliang Ren.
Machine translation Context-based approach Lucia Otoyo.
Evaluating the Contribution of EuroWordNet and Word Sense Disambiguation to Cross-Language Information Retrieval Paul Clough 1 and Mark Stevenson 2 Department.
Evaluation of the Statistical Machine Translation Service for Croatian-English Marija Brkić Department of Informatics, University of Rijeka
An Integrated Approach for Arabic-English Named Entity Translation Hany Hassan IBM Cairo Technology Development Center Jeffrey Sorensen IBM T.J. Watson.
Profile The METIS Approach Future Work Evaluation METIS II Architecture METIS II, the continuation of the successful assessment project METIS I, is an.
Advanced Signal Processing 05/06 Reinisch Bernhard Statistical Machine Translation Phrase Based Model.
Area Report Machine Translation Hervé Blanchon CLIPS-IMAG A Roadmap for Computational Linguistics COLING 2002 Post-Conference Workshop.
1 A Finite-State Approach to Machine Translation Srinivas Bangalore Giuseppe Riccardi AT&T Labs-Research NAACL 2001, Pittsburgh,
Péter Schönhofen – Ad Hoc Hungarian → English – CLEF Workshop 20 Sep 2007 Performing Cross-Language Retrieval with Wikipedia Participation report for Ad.
Recognition of spoken and spelled proper names Reporter : CHEN, TZAN HWEI Author :Michael Meyer, Hermann Hild.
Machine Translation  Machine translation is of one of the earliest uses of AI  Two approaches:  Traditional approach using grammars, rewrite rules,
Phrase Reordering for Statistical Machine Translation Based on Predicate-Argument Structure Mamoru Komachi, Yuji Matsumoto Nara Institute of Science and.
NUDT Machine Translation System for IWSLT2007 Presenter: Boxing Chen Authors: Wen-Han Chao & Zhou-Jun Li National University of Defense Technology, China.
Collocations and Information Management Applications Gregor Erbach Saarland University Saarbrücken.
Using Surface Syntactic Parser & Deviation from Randomness Jean-Pierre Chevallet IPAL I2R Gilles Sérasset CLIPS IMAG.
Indirect Supervision Protocols for Learning in Natural Language Processing II. Learning by Inventing Binary Labels This work is supported by DARPA funding.
What you have learned and how you can use it : Grammars and Lexicons Parts I-III.
For Monday Read chapter 24, sections 1-3 Homework: –Chapter 23, exercise 8.
For Friday Finish chapter 24 No written homework.
For Monday Read chapter 26 Last Homework –Chapter 23, exercise 7.
CSE467/567 Computational Linguistics Carl Alphonce Computer Science & Engineering University at Buffalo.
Improving Named Entity Translation Combining Phonetic and Semantic Similarities Fei Huang, Stephan Vogel, Alex Waibel Language Technologies Institute School.
Mutual bilingual terminology extraction Le An Ha*, Gabriela Fernandez**, Ruslan Mitkov*, Gloria Corpas*** * University of Wolverhampton ** Universidad.
Designing a Machine Translation Project Lori Levin and Alon Lavie Language Technologies Institute Carnegie Mellon University CATANAL Planning Meeting Barrow,
LREC 2004, 26 May 2004, Lisbon 1 Multimodal Multilingual Resources in the Subtitling Process S.Piperidis, I.Demiros, P.Prokopidis, P.Vanroose, A. Hoethker,
For Friday Finish chapter 23 Homework –Chapter 23, exercise 15.
Supertagging CMSC Natural Language Processing January 31, 2006.
Multi-level Bootstrapping for Extracting Parallel Sentence from a Quasi-Comparable Corpus Pascale Fung and Percy Cheung Human Language Technology Center,
1 Minimum Error Rate Training in Statistical Machine Translation Franz Josef Och Information Sciences Institute University of Southern California ACL 2003.
Example-based Machine Translation Pursuing Fully Structural NLP Sadao Kurohashi, Toshiaki Nakazawa, Kauffmann Alexis, Daisuke Kawahara University of Tokyo.
Generating Query Substitutions Alicia Wood. What is the problem to be solved?
Statistical Machine Translation Part II: Word Alignments and EM Alex Fraser Institute for Natural Language Processing University of Stuttgart
1 ICASSP Paper Survey Presenter: Chen Yi-Ting. 2 Improved Spoken Document Retrieval With Dynamic Key Term Lexicon and Probabilistic Latent Semantic Analysis.
For Monday Read chapter 26 Homework: –Chapter 23, exercises 8 and 9.
Overview of Statistical NLP IR Group Meeting March 7, 2006.
A Simple English-to-Punjabi Translation System By : Shailendra Singh.
Evaluating Translation Memory Software Francie Gow MA Translation, University of Ottawa Translator, Translation Bureau, Government of Canada
Review: Review: Translating without in-domain corpus: Machine translation post-editing with online learning techniques Antonio L. Lagarda, Daniel Ortiz-Martínez,
Learning to Generate Complex Morphology for Machine Translation Einat Minkov †, Kristina Toutanova* and Hisami Suzuki* *Microsoft Research † Carnegie Mellon.
LingWear Language Technology for the Information Warrior Alex Waibel, Lori Levin Alon Lavie, Robert Frederking Carnegie Mellon University.
Neural Machine Translation
Statistical Machine Translation Part II: Word Alignments and EM
--Mengxue Zhang, Qingyang Li
Presentation transcript:

Rapid Prototyping of Machine Translation Systems A Tale of Two Case Studies Srinivas Bangalore Giuseppe Riccardi AT&T Labs-Research Joint work with German Bordel and Vanessa Gaudin

Many thanks…to Alicia Abella Tirso Alonso Iker Arizmendi Barbara Hollister Mike Peñagarikano

Outline Machine Translation (MT) Past and Present Data Bottleneck and MT bootstrapping Consensus-based MT MT Evaluation Subjective and Objective Measures The Two Case Studies Demo

Machine Translation: Past and Present s present MT as code breaking, IBM-Georgetown Univ. demonstration Large bilingual dictionaries, linguistic and formal grammar motivated syntactic reordering, lots of funding, little progress ALPAC report: “there is no immediate or predictable prospect of useful machine translation” Translation continued in Canada, France and Germany. Beyond English- Russian translation. Meteo for translating weather reports. Systran in 1970 Emphasis on ‘indirect’ translation: semantic and knowledge-based. Advent of microcomputers. Translation companies: Systran, Logos, GlobalLink. Domain specific machine-aided translation systems. Corpus-based methods: IBM’s Candide, Japanese ‘example-based’ translation. Speech-to-Speech translation: Verbmobil, Janus. ‘Pure’ to practical MT for embedded applications: Cross-lingual IR

Corpus-based Translation Direct-translation methods relying on large parallel corpora. –Statistical Translation (IBM in early 90’s) stochastic generative model; parameters estimated for lexical choice, lexical reordering reordering based on string positions robust when encountered with new data –Example-based Translation (Japanese research) corpus of example translations match previous instances, retrieve closest match performs well for minor variants of previously encountered examples; typical in limited domains

Tree-based Alignment English: I’d like to charge this to my home phone Japanese: 私は これを 私の 家の 電話にチャージ したいのです Automatic algorithm (Alshawi, Bangalore and Douglas, 1998) 私は (I) これを (this) したいのです (like) チャージ (charge) 家の (home) 私の (my) 電話に (phone) I’d this like charge home my phone to

Statistical Translation Models Head Transducer Model (Alshawi, Bangalore and Douglas, 1998) –Context-free grammar based transduction model –Parsing complexity: O(n^6) Stochastic Finite-State Transducer Model (Bangalore and Riccardi 2000) –Approximation of context-free grammar based transduction model –Parsing complexity: O(n) –Tightly integrated with ASR

Translation in HMIHY

Bilingual Parallel Corpus Statistical translation techniques crucially depend on bilingual parallel corpus Typically, monolingual corpus is available How to create bilingual parallel corpus? Solution: Create bilingual parallel data with the help of translation houses + high quality translations -expensive and longer turn around time

Alleviating the Bilingual Data Bottleneck Creating Parallel Corpora: –Use of off-the-shelf translation engines (via the web) + Per sentence translation – No translation engine may be perfect; combine multiple translations Inducing Parallel Corpora: –Use of documents in multiple languages + Highly accurate translations + Unlimited data source – Document translations not sentence translations

Acquiring Bilingual Data Use of translation systems over the Web MT1 MT2MTn Web MT Interface Consensus Translation Monolingual Data Bilingual Data

Consensus Translation Translations differ in –Lexical choice –Word order Create consensus among different translations: –Multi-string alignment English: give me driving directions please MT1: deme direccionnes impulsoras por favor MT2: deme direccionnes por favor MT3: deme direccionnes conductores por favor MT4: deme las direccionnes que conducen satisfacen MT5: deme que las direccionnes tendencia a gradan

String Alignment Alignment of tokens between two strings –Insertion, deletion and substitution operations Two string alignment complexity: O(n^2) Multi-string alignment complexity: O(n^m) –Exponential in the number of strings (m) MT1: deme direccionnes impulsoras por favor MT2: deme direccionnes por favor Profile: * * d * *

Multi-String Alignment Progressive multi-sequence alignment (Feng and Doolittle 87) –Compute the edit distance and profiles for m*(m-1)/2 pairs –Repeat the following until one profile remains Construct profile strings for least edit distance string-string, string-profile or profile-profile pairs. Compute the edit distance between selected profile and the remaining strings and profiles

Multi-String Alignment S2 S1S3 P13 S4S5 P45 P1345 P12345 Strings to Align: S1, S2, S3, S4, S5

Consensus Translation (1) Result of multi-string alignment can be viewed as a “sausage” –Arcs represent words or phrases (possibly ) –Arcs between two states represent different translations of a word or phrase –Fan out at a states indicates disagreement in translation –Weights can be associated with each arc

Consensus Translation (2) Retrieving the consensus translation –Concatenate substrings from each segment of sausage –Majority vote: Substring with most number of votes from each segment of the lattice CMV = BestCostPath(Sausage) Some segments do not have a clear majority Use a posterior n-gram language model (λ) with weighting factor (α) CMV+LM = BestCostPath(Sausage o α*λ)

Consensus Translation (3) Retrieving a consensus translation English: give me driving directions please MT1: deme direccionnes impulsoras por favor MT2: deme direccionnes por favor MT3: deme direccionnes conductores por favor MT4: deme las direccionnes que conducen satisfacen MT5: deme que las direccionnes tendencia a gradan CT: deme direccionnes por favor

Outline Machine Translation (MT) Past and Present Data Bottleneck and MT bootstrapping Consensus-based MT MT Evaluation Subjective and Objective Measures The two Case Studies Demo

Spoken Dialog Corpus –Conference Registration System (“Innovation Forum”) –Average sentence length ~7 words/utt –Utterance from all dialog contexts Evaluation data –Small (~0.5K) (labeler agreement) –Large (~4K) (MT performance) Spoken Language Database

Criteria –Objective (string accuracy, parse accuracy) –Subjective (Labeler Annotation) Translator agreement (disagreement) –Not as straightforward as speech utterance transcriptions (ASR) –One-to-Many mapping (Language Generation) –Local phenomena ENGLISH Would you like to go out tomorrow night? ITALIAN Vuoi uscire domani sera? Vorresti uscire domani sera? Vuoi uscire fuori con me domani sera? Vuoi uscire con me domani sera? MT Evaluation

MT Evaluation (1) objective String alignment – no direct relation with semantics/syntax + objective + system incremental evaluation Test set of manual translation (300 sentences) String edit distance between reference string and result string (length in words: R) Translation String Accuracy = 1 – (M + I + D + S) / R

Evaluation Results (1) objective Translation accuracy CMV+LM51.0% CMV47.7% MT 129.8% MT 223.7% MT 335.2% MT 446.9% MT 549.7%

Semantic/Syntactic scale (1-3) 1 = The translation is semantically and syntactically correct 2 = The translation is semantically correct and the syntax has some flaws. 3 = The translation is neither semantically nor syntactically correct. Two Labelers The source language text was presented together with all hypotheses for the target language MT Evaluation (2) subjective

Labeler Score Distribution (Small test set) Labeler A Labeler B

Labeler Distributional Agreement CMV+LM0.02 MT system MT system MT system MT system MT system Binary random variable p_A(x=1) = 0.8 p_B(x=1) = 0.2 KL(p_A || p_B) ~ 1 Kulback-Leibler Distance

Evaluation (Large test set) Improve any MT system Decrease # Bad translation (score =3) Increase # Good translation ( score =1 or 2)

Hubbub…me (Instant Messaging)

Characteristics of Hubbub Data Human-human text-based interactions Open domain, conversations can be on any topic, may not be even task oriented Spontaneous chatty style of language (average 8 words per turn) Ungrammatical utterances and spelling errors Visual conversation context plays a crucial role in disambiguation Translation errors may be compensated based on the context of the conversation

Translation Accuracy Test set: 300 sentences

Summary Data Bottleneck solved by bootstrapping off existing MT systems Refine and Improve MT accuracy with Consensus-based MT Subjective and Objective Evaluation supports the improvement

Hubbub…me