Why Machine Translation? Machine Translation Overview & Introduction. Why Machine Translation? Why is translation hard?

Transcription

1 Why Machine Translation? Machine Translation Overview & Introduction Jörg Tiedemann Department of Linguistics and Philology Uppsala University translation is expensive on-line demand for translation (on-the-fly) globalization, growing export, lots of language pairs political issues (EU, support of minority languages...) tourism, movies, news... MT combines various aspects of computational linguistics Jörg Tiedemann 1/39 Jörg Tiedemann 2/39 Why Machine Translation? Why is translation hard? MT makes us laugh... Input: Hur många människor bor i Uppsala? Google 2009: How many people live in water? Input: Vem vann allsvenskan i fjol? Google 2010: Who stole headlines last year?... but constantly improves! Google 2011: How many people live in Uppsala? Who won the league last year? Languages are different on many levels: lexicon, syntax, semantics source language ambiguity cross-lingual divergences target language variation pragmatics, style, culture, background Computers have big problems with ambiguity. Jörg Tiedemann 3/39 Jörg Tiedemann 4/39

2 Translation by humans What can we do?... browsing quality A human translator needs... to understand the source language to know how to speak the target language (well!) knowledge about the relationship between source and target language knowledge about the topic of the text to be translated knowledge about culture, values, traditions and expectation av speakers of source and target language Computers have big problems with all of these issues Jörg Tiedemann 5/39 Jörg Tiedemann 6/39 What can we do? Outline Balance between quality and input restrictions Automatic translation depending on task general purpose semi-automatic sublanguage rough translation post-editing domain specific browsing quality editing quality publishing quality Motivation Course Overview fully automatic (gisting) computer-aided (human) translation (CAT) fully automatic (FAHQMT) What are the problems? MT history on-line service localization,... product manuals, limited domains Jörg Tiedemann 7/39 Jörg Tiedemann 8/39

3 General Course Related Information General Course Related Information Course: 5LN426 (7.5hp) introductory course presentation of main approaches (lectures) focus on practical sessions test & use some existing systems and tools develop & present projects Examination: pass obligatory lab assignments (G) translation project (graded) define & develop your own MT project present the project in class (2 seminars at the end) describe project in a report Jörg Tiedemann 9/39 Jörg Tiedemann 10/39 General Course Related Information What exactly is MT? Literature: Philipp Koehn: Statistical Machine Translation Cambridge University Press on-line material (scientific papers, slides,...) For the labs: you need an account on the STP machines... Website: joerg/mt11 Additional information might appear at any time during the course. MT = automatic translation from one language (source language) to another (target language) using computers MT translation memories and bilingual dictionaries MT - often: sentence-by-sentence translation MT often refers to translation of written text (not speech) semi-automatic: CAT = computer-aided translation Jörg Tiedemann 11/39 Jörg Tiedemann 12/39

4 Why is MT hard? Why is it hard to accept MT? Processing of languages is hard: analysis, information extraction, message understanding natural language generation, discourse fluency ambiguity resolution, synonymy & redundancy non-compositionality, idioms... domain detection & adaptation unrealistic expectations bad translations in available systems unexpected (not humanlike) errors problems of workflow integration Languages are different on many levels: lexicon, syntax, pragmatics style, culture, background changing now due to recent developments (!?) Jörg Tiedemann 13/39 Jörg Tiedemann 14/39 What are the problems? Source language ambiguity source language ambiguity target language variation cross-lingual divergences poäng point, points, credit,... var was, were (verb) each (pron) where (adv) every (adj) anta suppose admit (anta någon till...) Ambiguity is usually solved in context Jörg Tiedemann 15/39 Jörg Tiedemann 16/39

5 Lexical ambiguities across languages Language divergences and mismatches systematic differences between the 2 languages morphology (isolating vs polysynthetic, agglutinative vs fusion) syntax (SVO, SOV, VSO, argument structure, pro-drop) idiosyncratic and lexical differences differences in lexical ambiguity (polysemy) lexical gaps differences in tempus, aspect, voice different idiomatic/fixed expressions... (from Jurafsky & Martin, 2008: Speech and Language Processing) Jörg Tiedemann 17/39 Jörg Tiedemann 18/39 Divergences English - Spanish, sample of 19,000 sentences: 1 in 3 presents divergences (Dorr et al, 2002) Types (Habash, 2002) 1. categorical divergences: tener celos (N) (lit. to have jealousy ) to be jealous (A) 2. conflation: ir flotando (lit. to go floating ) to float 3. structural divergence: entrar en N (lit. to enter in N ) to enter N 4. head swapping: entrar corriendo (lit. to enter running ) to run in 5. thematic divergence: me gustan uvas (lit. to-me they-please grapes ) I like grapes Variation in Target language Redundancy of natural languages translate: Vid avslutad kurs... On completion of the course... After completion of the course... Having completed the course... After finishing the course... Once the course has been completed Which one is best? How do we decide that? Jörg Tiedemann 19/39 Jörg Tiedemann 20/39

6 Linguist s view on MT Engineer s view on MT Three steps for MT 1. source language analysis (understanding) 2. source to target language transfer (or interlingua) 3. target language generation (grammatical & fluent) appropriate tools? handling ambiguity? (transfer) formalism? interlingua? MT as decoding "I have a text in front of me which is written in Russian but I am going to pretend that it is really written in English and that it has been coded in some strange symbols. All I need to do is strip off the code in order to retrieve the information coded in the text" (Warren Weavers (1947/49), Rockefeller Foundation) (... and computers are good in encoding/decoding) Jörg Tiedemann 21/39 Jörg Tiedemann 22/39 Linguist s view on MT Engineer s view on MT What do we need for MT? What linguistic knowledge is needed to translate? 1. Phonological knowledge (speech-to-speech) 2. Morphological knowledge (inflection) 3. Syntactical knowledge (grammar) 4. Semantic knowledge (meaning) 5. Pragmatic knowledge (fluency, style, politeness,...) 1. more training data 2. statistical models 3. machine learning and efficient computing 4. more data! more data! more data! Fred Jelinek (IBM speech group) in the 80s: "Every time I fire a linguist the performance goes up" (... or something similar) Jörg Tiedemann 23/39 Jörg Tiedemann 24/39

7 Some words about the history of MT A brief history of MT (based on Hutchins) early days of computers: translation was thought to be one of the first tasks to be solved by computers (1930 s & 40s) MT = first task for language technology development of NLP/computational linguistics MT has a long history with many ups & downs 1. Precursors and pioneers, interest in cryptography, information theory, coding early computers, political/military interests 2. High expectations and disillusion, Georgetown experiment (rus-eng) big success three (classical) MT approaches (direct MT, interlingua-based, transfer-based) polarization: empiricists and theoretical (computational) linguists world-wide establishment of MT research groups problems: insufficient computer power, formal theories under development, no NLP tools/resources ALPAC report 1966: no need for MT (it s not going to work in near future) Jörg Tiedemann 25/39 Jörg Tiedemann 26/39 Misconceptions about MT A brief history of MT (based on Hutchins) "MT is a waste of time because you will never make a machine that can translate Shakespeare." "There was/is an MT system which translated The spirit is willing but the flesh is weak into the Russian equivalent of The vodka is good, but the steak is lousy. MT is useless". D.J. Arnold, Lorna Balkan, Siety Meijer, R.Lee Humphreys and Louisa Sadler Machine Translation: an Introductory Guide, 1994 The quiet decade, domain-specific translation (controlled language) mainly interlingua approaches but no real progress Operational and commercial systems, Systran, METAL computer-aided translation (especially in Japan) Revival of MT research, Data-driven MT, 1990 SMT, EBMT, synchronous grammars, hybrid models Jörg Tiedemann 27/39 Jörg Tiedemann 28/39

8 Where are we now? Two main paradigms Today: more automatic translations than manual ca 50 million translation requests every day on the Internet (van der Meer, 2008) Google: provides on-line translation for 63 languages approaches inspired by linguistic theory data-driven techniques (re-using existing translations) Let s look at the classical rule-based systems first Tasks: higher quality, other domains, other language pairs Jörg Tiedemann 29/39 Jörg Tiedemann 30/39 The Vauquois triangle (translation triangle) The Vauquois triangle: Classical MT approaches some intermediate symbolic representation (from Jurafsky & Martin, 2008: Speech and Language Processing) Jörg Tiedemann 31/39 Jörg Tiedemann 32/39

9 Direct Translation Transfer-based MT morphological analysis word-by-word using large bilingual dictionaries (may include phrases) local reordering using simple rules morphological generation Example (from Jurafsky & Martin) Input: Mary didn t slap the green witch After 1: Morphology Mary DO-PAST not slap the green witch After 2: Lexical Transfer Maria PAST no dar una bofetada a la verde bruja After 3: Local reordering Maria no dar PAST una bofetada a la bruja verde After 4: Morphology Maria no dió una bofetada a la bruja verde 3 steps: 1. source language analysis 2. transfer (source to target structures) 3. target language generation transfer rules: lexical transfer syntactic transfer semantic transfer Jörg Tiedemann 33/39 Jörg Tiedemann 34/39 Transfer-based MT Transfer rules to cover systematic structural differences Informal description of some transformations English to Spanish: NP Adjective1 Noun2 NP Noun2 Adjective1 Chinese to English: VP PP[+Goal] V VP V PP[+Goal] English to Japanese: VP V NP VP NP V NP NP1 RelClause2 NP RelClause2 NP1 Source language analysis depends on features necessary for transfer (purely syntactic, semantic features,...) Lexical transfer rules (phrasal) bilingual dictionaries Jörg Tiedemann 35/39 Interlingua-based MT Basic idea: use meaning representation semantic analysis of source language input (formal) representation of meaning using interlingua ( no language-pair specific transfer necessary!) target language generation from meaning representation Nice idea, but... Where are the tools? What is a good way of representing meaning? Many problems with ambiguity and language divergences Jörg Tiedemann 36/39

10 Interlingua-based MT Summary Example interlingual representation: Mary did not slap the green witch long, bumpy history rule-based systems vs. data-driven systems linguistic theory vs. examples & statistics large demand for MT world-wide MT quality depending on task MT = hot topic again Jörg Tiedemann 37/39 Jörg Tiedemann 38/39 Next lecture: more on rule-based systems domain-specific approaches evaluation of MT output (manual & automatic) first lab session: evaluation of MT output a little experiment with understandability Jörg Tiedemann 39/39