Fall 2005

Lecture Notes #9

EECS 595 / LING 541 / SI 661

Natural Language Processing

Machine Translation

Example (from the Hansards corpus)

• English• <s id=960001> I would like the government and the Postmaster General to

agree that we place the union and the Postmaster General under trusteeship so that we can look at his books and records, including those of his management people and all the memos he has received from them, some of which must have shocked him rigid.

• <s id=960002> If the minister would like to propose that, I for one would be prepared to support him.

• French• <s id=960001> Je voudrais que le gouvernement et le ministre des Postes

conviennent de placer le syndicat et le ministre des Postes sous tutelle afin que nous puissions examiner ses livres et ses dossiers, y compris ceux de ses collaborateurs, et tous les mémoires qu'il a reçus d'eux, dont certains l'ont sidéré.

• <s id=960002> Si le ministre voulait proposer cela, je serais pour ma part disposé à l'appuyer.

Example

• These lies are like their father that begets them; gross as a mountain, open, palpable(Henry IV, Part 1, act 2, scene 2)

Language similarities and differences

• Word order (SVO: English, Mandarin, VSO: Irish, Classical Arabic, SOV: Hindi, Japanese)

• Prepositions (Jap.) (to Mariko, Mariko-ni)• Lexical distinctions (Sp.):

– the bottle floated out

– la botella salió flotando

• Brother (Jap.) = otooto (younger), oniisan (older)• They (Fr.) = elles (feminine), ils (masculine)

Acrobat Document

Why is Machine Translation Hard?

• Analysis• Transfer/interlingua• Generation

INPUT OUTPUT2OUTPUT2OUTPUT2

OUTPUT1

OUTPUT3

Basic Strategies of MT

• Direct Approach– 50’s,60’s – naïve

• Indirect: Interlingua– No looking back– Language-neutral– No influence on the target language

• Indirect: Transfer– Preferred

F E

I

Levels of Linguistic Processing

• Phonology

• Orthography

• Morphology (inflectional, derivational)

• Syntax (e.g., agreement)

• Semantics (e.g., concrete vs. abstract terms)

• Discourse (e.g., use of pronouns)

• Pragmatics (world knowledge)

Category Ambiguity

• Morphological ambiguity (“Wachtraum”)

• Part-of-speech (category) ambiguity (e.g. “round”)

• Some help comes from morphology (“rounding”)

• Using syntax, some ambiguities disappear (context dictates category)

Homography and Polysemy

• Homographs: (“light”, “club”, “bank”)

• Polysemous words: (“channel”, “crane”)

• for different categories - syntax

• for same category - semantics

Structural Ambiguity

• Humans can have multiple interpretations (parses) for the same sentence

• Example: prepositional phrase attachment

• Use context to disambiguate

• For machine translation, context can be hard to define

Use of Linguistic Knowledge

• Subcategorization frames

• Semantic features (is an object “readable”?)

Contextual Knowledge

• In practice, very few sentences are truly ambiguous

• Context makes sense for humans (“telescope” example), not for machines

• no clear definition of context

Other Strategies

• Pick most natural interpretation

• Ask the author

• Make a guess

• Hope for a free ride

• Direct transfer

Acrobat DocumentAcrobat DocumentAcrobat Document

Anaphora Resolution

• Use of pronouns (“it”, “him”, “himself”, “her”)

• Definite anaphora (“the young man”)

• Antecedents

• Same problems as for ambiguity resolution

• Similar solutions (e.g., subcategorization)

The Noisy Channel Model

• Source-channel model of communication

• Parametric probabilistic models of language and translation

• Training such models

Statistics

• Given f, guess e

ef

e’E F F E

encoder decoder

e’ = argmax P(e|f) = argmax P(f|e) P(e)e e

translation model language model

Parametric probabilistic models

• Language model (LM)

• Deleted interpolation

• Translation model (TM)

P(e) = P(e1, e2, …, eL) = P(e1) P(e2|e1) … P(eL|e1 … eL-1)

P(eL|e1 … eK-1) P(eL|eL-2, eL-1)

Alignment: P(f,a|e)

English and Cebuano

In the beginning God created the heaven and the earth.Sa sinugdan gibuhat sa Dios ang mga langit ug ang yuta.

And God called the firmament Heaven. Ug gihinganlan sa Dios ang hawan nga Langit.

And God called the dry land Earth Ug ang mamala nga dapit gihinganlan sa Dios nga Yuta

• use: co-occurrence, word order, cognates• corpora are needed • sentence alignment needs to be done first

Statistical MT

Statistical MTTranslate from French: “une fleur rouge”?

p(e) p(f|e) p(e)*p(f|e)

a flower red low high low

red flower a low high low

flower red a low high low

a red dog high low low

dog cat mouse low low low

a red flower high high high

Issues to deal with

• word order:– I like to drink coffee– watashi wa kohii o nomu no ga suki desu– I-subj coffee-obj drink-dat-rheme like

• vocabulary:– wall– pared, muro

• phrases:– play– pièce de théâtre

MT/noisy channel models

• Text-to-text (summ), also text-to-signal, speech recognition, OCR, spelling correction

• P(text|pixels) = P(text) P(pixels|text)

IBM’s EM trained models (1-5)

• Word translation

• Local alignment

• Fertilities

• Class-based alignment

• Non-deficient algorithm (avoid overlaps, overflow)

Steps

• Tokenization

• Sentence alignment (1-1, 2-2, 2-1 mappings)– Church and Gale (based on sentence length)– Church (sequences of 4-grams) – based on

cognates– Melamed (longest common subsequence of

words) – also cognates

Model 1

• Alignments– La maison bleue– The blue house– Alignments: {1,2,3}, {1,3,2}, {1,3,3}, {1,1,1}– All are equally likely

• Conditional probabilities– P(f|A,e) = ?

Model 1 (cont’d)

• Algorithm– Pick length of translation

– Choose an alignment

– Pick the French words

– That gives you P(f,A|e)

– We need P(f|A,e)

– Use EM (expectation-maximisation) to find the hidden variables

– (see Kevin Knight’s tutorial)

Model 1

• We need p(f|e) but we don’t know the word alignments (which are assumed to be equally likely)

m

jajm jefp

l

ceAfpeApeAfp

1

)|()1(

),|(*)|()|,(

Model 2

• Distortion parameters D(i|j,l,m) – i and j are words in the two sentences– l and m are the lengths of these sentences.

Model 3

• Fertility

• P(i|e)

• Examples– (a) play = pièce de théâtre– (to) place = mettre en place

• p1 is an extra parameter that defines 0

Current work

• Handling phrases

• Using syntax– In the model– In discriminative reranking

• Low density languages

Evaluation

• Human judgements: adequacy, grammaticality

• Automatic methods– BLEU– ROUGE

When does MT work?

• Machine-Aided Translation (MAT)

• Restricted Domains (e.g., technical manuals)

• Restricted Languages (sublanguages)

• To give the reader an idea of what the text is about

Dialogueand conversational agents

REMEMBER TO READ THE NEW VERSION OF THIS CHAPTER ON THE WEB!

Abbott You know, strange as it may seem, they give ball players nowadays very peculiar names...Now, on the Cooperstown team we have Who's on first, What's on second, I Don't Know is on third-

Costello That's what I want to find out. I want you to tell me the names of the fellows on the Cooperstown team.

Abbott I'm telling you. Who's on first, What's on second, I Don't Know is on third.

Costello You know the fellows' names?

Abbott Yes.

Costello Well, then, who's playin' first?

Abbott Yes.

Costello I mean the fellow's name on first base.

Abbott Who.

Costello The fellow's name on first base for Cooperstown.

Abbott Who.

Costello The guy on first base.

Abbott Who is on first base.

Costello Well, what are you asking me for?

Abbott I'm not asking you--I'm telling you. Who is on first.

Costello I'm asking you--who's on first?

Abbott That's the man's name.

Costello That's who's name?

Abbott Yes.

Costello Well, go ahead, tell me!

Abbott Who.

Costello The guy on first.

Abbott Who.

Costello The first baseman.

Abbott Who is on first.

Costello Have you got a first baseman on first?

Abbott Certainly.

Costello Well, all I'm trying to find out is what's the guy's name on first base.

Abbott Oh, no, no. What is on second base.

Costello I'm not asking you who's on second.

What makes dialogue different

• Turns and utterances (turn-taking)• Turn-taking rules

– At each TRP (transition-relevance place):• designated speaker, any speaker, current speaker

– Barge-in possible

• Significant silence– A: Is there something bothering you or not? (1.0 s)– A: Yes or no? (1.5 s)– A: Eh?– B: No.

fig19.01.pdf

Grounding

• Common ground between speaker and hearer.• A: … returning on flight 1118• C: mm hmmm (backchannel, acknowledgment token)• Other continuers:

– Continued attention– Relevant next contribution– Acknowledgement (e.g. “sure”)– Demonstration (paraphrasing, reformulating)– Display (repeat verbatim)

• Example:• C: I will take the 5 pm flight on the 11th.• A: On the 11th?

Conversational Implicature

• Example:– When do you want to travel?– I have a meeting there early in the morning on

the 13th.

• Implicature: licensed inferences reasonable hearers can make.

• Quantity:– Agent: “there are three non-stop flights daily”

Grice’s maxims

• Maxim of quantity– make your contribution informative

– but not more than needed

• Maxim of quality– do not say what you believe is false

– do not say that for which you lack evidence

• Maxim of relevance

• Maxim of manner– avoid ambiguity

– avoid obscurity

– be brief

– be orderly

Dialogue acts

• Performative sentences:– I name this ship the Titanic– I second that motion– I bet you five dollars that it will snow tomorrow

• Speech acts:– locutionary acts: uttering a sentence with a particular

meaning– illocutionary acts: asking, promising, answering…– perlocutionary acts: producing effects upon the feelings,

thoughts, or actions of the addressee

Speech acts (cont’d)

• Assertives: suggesting, putting forward, swearing, boasting, concluding

• Directives: asking, ordering, requesting, inviting, advising, begging

• Commissives: promising, planning, vowing, betting, opposing

• Expressives: thanking, apologizing, welcoming, deploring

• Declarations: I resign, you’re fired.

• DAMSL - Dialogue Act Markup in Several Layers

• Agreement (Accept, Maybe, Reject-Part, Hold)

• Answer

• Understanding (Signal-not-understood, Signal-understood, ack, repeat-rephrase, completion)

Automatic interpretation of dialogue acts

fig19.02.pdf

Techniques for DA recognition

• Plan theoretic (agents, assumptions, goals)

• Cue-based (“please”, “are you?”, rising pitch, stress - agreement vs. backchannel)

• Statistical approaches

fig19.03.pdffig19.04.pdf