CS460/626 : Natural Language Processing/Speech, NLP and the Web

(Lecture 31 Introduction to Machine Translation)

Pushpak BhattacharyyaCSE Dept., IIT Bombay

29th March, 2012

Introduction to MT Machine Translation (MT) is a technique to translate

texts from one natural language to another natural language using a machine

Translated text should have two desired properties:

Adequacy: Meaning should be conveyed correctly

Fluency: Text should be fluent in the target Fluency: Text should be fluent in the target language

Translation between distant languages is a difficult task

Handling Language Divergence is a major challenge

2

Kinds of MT Systems(point of entry from source to the target text)

Deep understanding level

Interlingual level

Ascendin

g transfe

rLogico-semantic level

Conceptual transfer

Semantic transfer

Ontological interlingua

Semantico-linguistic interlingua

SPA-structures (semantic& predicate-argument)

Ascendin

g transfe

r

Syntactico-functional level

Morpho-syntactic level

Syntagmatic level

Graphemic level Direct translation

Syntactic transfer (surface)

Syntactic transfer (deep)

Multilevel transfer

F-structures (functional)

C-structures (constituent)

Tagged text

Text

Mixing levels Multilevel description

Semi-direct translation Descending transfers

3

Why is MT difficult: Language Divergence

One of the main complexities of MT: Language Divergence

Languages have different ways of Languages have different ways of expressing meaning

Lexico-Semantic Divergence

Structural Divergence

4

English-IL Language Divergence with illustrations from Hindi(Dave, Parikh, Bhattacharyya, Journal of MT, 2002)

Language Divergence Theory: Lexico-Semantic Divergences

Conflational divergence F: vomir; E: to be sick E: stab; H: churaa se maaranaa (knife-with hit) S: Utrymningsplan; E: escape plan

Structural divergence Structural divergence E: SVO; H: SOV

Categorial divergence Change is in POS category

Not communicative (AdjVerb)

5

Language Divergence Theory: Lexico-Semantic Divergences (cntd)

Head swapping divergence E: Prime Minister of India; H: bhaarat ke pradhaanmantrii (India-of Prime Minister)

Lexical divergence E: advise; H: paraamarsh denaa (advice give): NounIncorporation- very common Indian LanguageIncorporation- very common Indian LanguagePhenomenon

6

Language Divergence Theory: Syntactic Divergences

Constituent Order divergence

E: Singh, the PM of India, will address the nationtoday; H: bhaarat ke pradhaan mantrii, singh, (India-of PM, Singh)

Adjunction Divergence

E: She will visit here in the summer; H: vah yahaagarmii meM aayegii (she here summer-in will come)

E: She will visit here in the summer; H: vah yahaagarmii meM aayegii (she here summer-in will come)

Preposition-Stranding divergence

E: Who do you want to go with?; H: kisake saath aapjaanaa chaahate ho? (who with)

7

Language Divergence Theory: Syntactic Divergences

Null Subject Divergence

E: I will go; H: jaauMgaa (subject dropped)

Pleonastic Divergence

E: It is raining; H: baarish ho rahii haai (rainhappening is: no translation of it)happening is: no translation of it)

8

Language Divergence exists even for close cousins(Marathi-Hindi-English: case marking and postpositions do not transfer easily)

(immediate past) ? ! (M) ? (H) When did you come? Just now (I came) (H) kakhan ele? ei elaam/eshchhi (B)

(certainty in future) !

(certainty in future) ! (M) ! (H) He is in for a thrashing. (E) ekhan o maar khaabei/khaachhei

(assurance) ! . (M) $ (H) I will see you tomorrow. (E) aami kaal tomaar saathe dekhaa korbo/korchhi (B)

9

Can MT make a difference

Yes: MT is ubiquitoushttp://www.cfilt.iitb.ac.in/judicial/

http://www.cfilt.iitb.ac.in:11080/XeroxIITBSMT/Translate

http://translate.google.com

Helping a person in distress (1/3)

French Lady (FL): (in French, translated into English by Google translate) Do you know how far it is to the Bangalore you know how far it is to the Bangalore main bus stand from the airport?

Self: (in English, translated into French by GT) It is about 10 KM.

FL: Can I take a bus?

Helping a person in distress (2/3)

FL: I think that is expensive.

Self: Yes about 3000 Rupees, about 50 Self: Yes about 3000 Rupees, about 50 Euros.

FL: No, I will take a bus.

Self: Did you hurt your ankle. Do you need medical help?

Helping a person in distress (3/3)

Self: Should I look for a doctor?

FL: No, thank you, I will consult in FL: No, thank you, I will consult in Mysore.

All this conversation occurred via Google translation. The lady and I took turns to type in the sentences in respective translation boxes.

Translation is Ubiquitous

Between Languages

Delhi is the capital of India

Between dialects

Example next slide

Between registers

My mom not well.

My mother is unwell (in a leave application)

Between dialects (1/3)

Lage Raho Munnabhai: an excellent example

Scene: Munnabhai (Sanjay Dutt) is Prof. MurliPrasad Sharma being interviewed with some citizens asking questions in presence of Jahnavi(Vidya Baalan)(Vidya Baalan)

Question by citizen: , . $ .

Between dialects (2/3) Bapu from behind invisible to others:

Munnabhai

#

Bapu

% Munnabhai

full country % Bapu

* Munnabhai

,%-

Between dialects (3/3) Bapu

$ Munnabhai

( $ ,

Bapu

$ Munnabhai

/ 0 $ , , /, heart heart !

Interlingua Based MT

18

MT: EnConverion + Deconversion

EnglishHindi

Interlingua(UNL)

FrenchChinese

generation

Analysis

19

Challenges of interlingua generation

Mother of NLP problems - Extract meaning from a sentence!

Almost all NLP problems are sub-problems

Named Entity Recognition Named Entity Recognition

POS tagging

Chunking

Parsing

Word Sense Disambiguation

Multiword identification

and the list goes on...

UNL: a United Nations project

Started in 1996 10 year program 15 research groups across continents First goal: generatorsFirst goal: generators Next goal: analysers (needs solving various ambiguity

problems) Current active language groups

UNL_French (GETA-CLIPS, IMAG) UNL_Hindi (IIT Bombay with additional work on UNL_English)

UNL_Italian (Univ. of Pisa) UNL_Portugese (Univ of Sao Paolo, Brazil) UNL_Russian (Institute of Linguistics, Moscow) UNL_Spanish (UPM, Madrid)

21

World-wide Universal Networking Language (UNL) Project

UNL

English Russian

Marathi

22

Japanese

Hindi

Spanish

Language independent meaning representation.

Others

UNL represents knowledge: John eats rice with a spoon

Semantic relations

attributes

Universal words

Repositoryof 42SemanticRelations

and84 attributelabels

23

Hindi Generation from Interlingua (UNL)

(Joint work with S. Singh, M. Dalal, V. Vachhani, Om Damani

MT Summit 2007)

HinD ArchitectureDeconversion = Transfer + Generation

Step-through the Deconverter Module Output

UNL Expression

obj(contact(

Lexeme Selection

contact farmer this you region taluka manchar khatav

Case Identification

* * *

contact farmer* this you region* taluka* manchar

you

this

farmer

agtobj

pur

plc

contact

orregion

nam

:01

khatav Morphology Generation

contact .@imperative farmer.@pl this you region

taluka manchar Khatav

Function Word Insertion

contact farmers this for you region

or taluka of Manchar Khatav

Linearization This for you manchar region or khatav

| taluka of farmers contact

nam

or

khatav

manchar

taluka

nam

Lexeme Selection[ ]{}"contact(icl>communicate(agt>person,obj>person)) (V,VOA,VOA-

ACT,VOA-COMM,VLTN,TMP,CJNCT,N-V,link,Va)

[ ]{}"contact(icl>representative)

(N,ANIMT,FAUNA,MML,PRSN,Na)

Lexical Choice is unambiguousLexical Choice is unambiguous

you

this farmer

agtobj contact

pur

Case Marking

Relation Parent+ Parent- Child+ Child- Obj V VINT N Agt V @past N

Depends on UNL Relation and the properties of the nodes Depends on UNL Relation and the properties of the nodes Case get transferred from head to modifiers

you

this *farmer

agtobj contact

pur

Morphology Generation: NounsThe boy saw me.

Boys saw me.

The King saw me.

Kings saw me.

Suffix Attribute values

uoM @N,@NU,@M,@pl,@oblique

U @N,@NU,@M,@sg,@oblique I @N,@NI,@F,@sg,@oblique iyoM @N,@NI,@F,@pl,@oblique

oM @N,@NA,@NOTCH,@F,@pl,@oblique

Kings saw me.

Verb MorphologySuffix Tense Aspect Mood N Gen P V

E

-e rahaa thaa

@past @progress - @sg @male 3rd e

-taa hai @present @custom - @sg @male 3rd -

-iyaa thaa @past @complete - @sg @male 3rd I

saktii hain @present - @ability @pl @female 3rd A

After Morphology Generation

you

this

farmer

agtobj contact

pur

you

this

farmer

agtobjcontact

pur

Function Word Insertion

Rel Par+ Par- Chi+ Chi- Ch/FW

Obj V VINT N#ANIMT @topic

you

this

farmer

agtobj contact

pur

Obj V VINT N#ANIMT @topic

Linearization

This you manchar region

you

this

farmer

agtobj

pur

plc

Contact

khatav taluka farmers

contact

nam

orregion

khatav

manchar

taluka

nam

:01

Syntax Planning: Assumptions The relative word order of a

UNL relations relata does not depend on:

Semantic Independence: the semantic properties of the this

contact

pur

semantic properties of the relata.

Context Independence: the rest of the expression.

The relative word order of various relations sharing a relatum does not depend on

Local Ordering: the rest of the expression.

this

you

this

farmer

agtobj

contact

pur

you

this

farmer

agtobj contact

pur

Syntax Planning: Strategy Divide a nodes relations in

Before_set = {obj,pur,agt}

After_set = {}

agt aoj obj ins

Agt L L L

Aoj L L

Obj R

Ins

obj pur

agt

Topo Sort each group: pur agt objthis you farmer

Final order: this you farmer contact

region

khatav

manch

taluka

Syntax Planning AlgoStack Before

Current

After

Current

Output

region

farmer

contact

this

you

manchar taluka

you

this

farmer

agtobj

pur

plc

Contact

manchar taluka

manchar

region

taluka

farmer

contact

region

taluka

farmer

Contact

this

you

manchar

taluka

farmer

contact

this

you

manchar

region

plc

nam

orregion

khatav

manchar

taluka

nam

:01

All Together (UNL -> Hindi) Module Output

UNL Expression

obj(contact(..

Lexeme Selection

contact farmer this you region taluka manchar khatav

Case Identification

* * *

contact farmer* this you region* taluka* manchar Identification

contact farmer* this you region* taluka* manchar

khatav Morphology Generation

contact .@imperative farmer.@pl this you region

taluka manchar Khatav

Function Word Insertion

contact farmers this for you region

or taluka of Manchar Khatav

Syntax Linearization

This for you manchar region or khatav

| taluka of farmers contact

How to Evaluate UNL Deconversion

UNL -> Hindi

Reference Translation needs to be generated from UNL

Needs expertise Needs expertise

Compromise: Generate reference translation from original English sentence from which UNL was generated

Works if you assume that UNL generation was perfect

Note that fidelity is not an issue

Manual Evaluation Guidelines

Fluency of the given translation is:(4) Perfect: Good grammar(3) Fair: Easy-to-understand but flawed grammar(2) Acceptable: Broken - understandable with effort(1) Nonsense: Incomprehensible(1) Nonsense: Incomprehensible

Adequacy: How much meaning of the reference sentence isconveyed in the translation:(4) All: No loss of meaning(3) Most: Most of the meaning is conveyed(2) Some: Some of the meaning is conveyed(1) None: Hardly any meaning is conveyed

Sample Translations

Hindi Output Fluency Adequacy

0.5 ! 10 )

2 3

,- / 0 2 34 3 4,- / 0 2 34 6 2 8 ,

3 4

8 3< 3<

1 1

)> 8 4 4

- B D

2 3

Results BLEU Fluency Adequacy

Geometric Average 0.34 2.54 2.84 Arithmetic Average 0.41 2.71 3.00 Standard Deviation 0.25 0.89 0.89 Pearson Cor. BLEU 1.00 0.59 0.50

Pearson Cor. Fluency 0.59 1.00 0.68

Fluency vs Adequacy

315 3 4

34

165

2662

155138

210 12

131

168

0

50

100

150

200

1 2 3 4

Fluency

Nu

mb

er o

f s

ente

nce

s

Adequacy 1 Adequacy 2 Adequacy 3 Adequacy 4

Good Correlation between Fluency and BLUE Strong Correlation betweenFluency and Adequacy Can do large scale evaluation using Fluency alone

Caution: Domain diversity, Speaker diversity

Summary: interlingua based MT

English to Hindi

Rule governed

High level of linguistic expertise needed High level of linguistic expertise needed

Takes a long time to build (since 1996)

But produces great insight, resources and tools

42