cs60057 speech &natural language processing

Lecture 1, 7/21/2005 Natural Language Processing 1

CS60057Speech &Natural Language

Processing

Autumn 2007

Lecture 2

26 July 2007


Why is NLP difficult?

Because Natural Language is highly ambiguous. Syntactic ambiguity

The president spoke to the nation about the problem of drug use in the schools from one coast to the other.

has 720 parses. Ex:

“to the other” can attach to any of the previous NPs (ex. “the problem”), or the head verb 6 places

“from one coast” has 5 places to attach …


Why is NLP difficult? Word category ambiguity

book --> verb? or noun? Word sense ambiguity

bank --> financial institution? building? or river side? Words can mean more than their sum of parts

make up a story Fictitious worlds

People on mars can fly. Defining scope

People like ice-cream. Does this mean that all (or some?) people like ice cream?

Language is changing and evolving I’ll email you my answer. This new S.U.V. has a compartment for your mobile phone. Googling, …


Dealing with Ambiguity

Four possible approaches: Tightly coupled interaction among processing

levels; knowledge from other levels can help decide among choices at ambiguous levels.

Pipeline processing that ignores ambiguity as it occurs and hopes that other levels can eliminate incorrect structures.


Resolve Ambiguities We will introduce models and algorithms to resolve ambiguities at

different levels. part-of-speech tagging -- Deciding whether duck is verb or noun. word-sense disambiguation -- Deciding whether make is create or cook.

lexical disambiguation -- Resolution of part-of-speech and word-sense ambiguities are two important kinds of lexical disambiguation.

syntactic ambiguity -- her duck is an example of syntactic ambiguity, and can be addressed by probabilistic parsing.


Resolve Ambiguities (cont.)I made her duck

S S

NP VP NP VP

I V NP NP I V NP

made her duck made DET N

her duck


Dealing with Ambiguity

Three approaches: Tightly coupled interaction among processing

levels; knowledge from other levels can help decide among choices at ambiguous levels.

Pipeline processing that ignores ambiguity as it occurs and hopes that other levels can eliminate incorrect structures.

Syntax proposes/semantics disposes approach Probabilistic approaches based on making the

most likely choices


Models and Algorithms

By models I mean the formalisms that are used to capture the various kinds of linguistic knowledge we need.

Algorithms are then used to manipulate the knowledge representations needed to tackle the task at hand.


Models to Represent Linguistic Knowledge

Different formalisms (models) are used to represent the required linguistic knowledge.

State Machines -- FSAs, HMMs, ATNs, RTNs Formal Rule Systems -- Context Free Grammars,

Unification Grammars, Probabilistic CFGs. Logic-based Formalisms -- first order predicate logic,

some higher order logic. Models of Uncertainty -- Bayesian probability theory.


Algorithms

Many of the algorithms that we’ll study will turn out to be transducers; algorithms that take one kind of structure as input and output another.

Unfortunately, ambiguity makes this process difficult. This leads us to employ algorithms that are designed to handle ambiguity of various kinds


Algorithms

In particular.. State-space search

To manage the problem of making choices during processing when we lack the information needed to make the right choice

Dynamic programming To avoid having to redo work during the course of a

state-space search CKY, Earley, Minimum Edit Distance, Viterbi,

Baum-Welch


State Space Search

States represent pairings of partially processed inputs with partially constructed representations.

Goals are inputs paired with completed representations that satisfy some criteria.

As with most interesting problems the spaces are normally too large to exhaustively explore. We need heuristics to guide the search Criteria to trim the space


Dynamic Programming

Don’t do the same work over and over. Avoid this by building and making use of solutions to

sub-problems that must be invariant across all parts of the space.


Languages

Languages: 39,000 languages and dialects (22,000 dialects in India alone)

Top languages: Chinese/Mandarin (885M), Spanish (332M), English (322M), Bengali (189M), Hindi (182M), Portuguese (170M), Russian (170M), Japanese (125M)

Source: www.sil.org/ethnologue, www.nytimes.com Internet: English (128M), Japanese (19.7M), German (14M), Spanish

(9.4M), French (9.3M), Chinese (7.0M) Usage: English (1999-54%, 2001-51%, 2003-46%, 2005-43%) Source: www.computereconomics.com


The Description of Language

Language = Words and Rules Dictionary (vocabulary) + Grammar Dictionary set of words defined in the language. open (dynamic)

Traditional - paper based Electronic - machine readable dictionaries; can be obtained from paper-

based Grammar set of rules which describe what is allowable in a language Classic Grammars

meant for humans who know the language definitions and rules are mainly supported by examples no (or almost no) formal description tools; cannot be programmed

Explicit Grammar (CFG, Dependency Grammars, Link Grammars,...) formal description can be programmed & tested on data (texts)


Levels of (Formal) Description 6 basic levels (more or less explicitly present in most theories):

and beyond (pragmatics/logic/...)

meaning (semantics)

(surface) syntax

morphology

phonology

phonetics/orthography

Each level has an input and output representation output from one level is the input to the next (upper) level sometimes levels might be skipped (merged) or split


Phonetics/Orthography Input:

acoustic signal (phonetics) / text (orthography)

Output: phonetic alphabet (phonetics) / text (orthography)

Deals with: Phonetics:

consonant & vowel (& others) formation in the vocal tract classification of consonants, vowels, ... in relation to frequencies,

shape & position of the tongue and various muscles intonation

Orthography: normalization, punctuation, etc.


Phonology Input:

sequence of phones/sounds (in a phonetic alphabet); or “normalized” text (sequence of (surface) letters in one language’s alphabet) [NB: phones vs. phonemes]

Output: sequence of phonemes (~ (lexical) letters; in an abstract alphabet)

Deals with: relation between sounds and phonemes (units which might have

some function on the upper level) e.g.: [u] ~ oo (as in book), [æ] ~ a (cat); i ~ y (flies)


Morphology

Input: sequence of phonemes (~ (lexical) letters)

Output: sequence of pairs (lemma, (morphological) tag)

Deals with: composition of phonemes into word forms and their

underlying lemmas (lexical units) + morphological categories (inflection, derivation, compounding)

e.g. quotations ~ quote/V + -ation(der.V->N) + NNS.


(Surface) Syntax Input:

sequence of pairs (lemma, (morphological) tag) Output:

sentence structure (tree) with annotated nodes (all lemmas, (morphosyntactic) tags, functions), of various forms

Deals with: the relation between lemmas & morphological categories and

the sentence structure uses syntactic categories such as Subject, Verb, Object,... e.g.: I/PP1 see/VB a/DT dog/NN ~ ((I/sg)SB ((see/pres)V (a/ind dog/sg)OBJ)VP)S


Meaning (semantics) Input:

sentence structure (tree) with annotated nodes (lemmas, (morphosyntactic) tags, surface functions)

Output: sentence structure (tree) with annotated nodes (semantic

lemmas, (morpho-syntactic) tags, deep functions) Deals with:

relation between categories such as “Subject”, “Object” and (deep) categories such as “Agent”, “Effect”; adds other cat’s

e.g. ((I)SB ((was seen)V (by Tom)OBJ)VP)S ~ (I/Sg/Pat/t (see/Perf/Pred/t) Tom/Sg/Ag/f)


...and Beyond

Input: sentence structure (tree): annotated nodes (autosemantic

lemmas, (morphosyntactic) tags, deep functions) Output:

logical form, which can be evaluated (true/false) Deals with:

assignment of objects from the real world to the nodes of the sentence structure

e.g.: (I/Sg/Pat/t (see/Perf/Pred/t) Tom/Sg/Ag/f) ~ see(Mark-Twain[SSN:...],Tom-Sawyer[SSN:...])[Time:bef 99/9/27/14:15][Place:39ş19’40”N76ş37’10”W]


Three Views

Three equivalent formal ways to look at what we’re up to (not including tables)

Regular Expressions

Regular LanguagesFinite State Automata


Transition

Finite-state methods are particularly useful in dealing with a lexicon.

Lots of devices, some with limited memory, need access to big lists of words.

So we’ll switch to talking about some facts about words and then come back to computational methods


MORPHOLOGY


Morphology Morphology is the study of the ways that words are built up from

smaller meaningful units called morphemes (morph = shape, logos = word)

We can usefully divide morphemes into two classes Stems: The core meaning bearing units Affixes: Bits and pieces that adhere to stems to change their

meanings and grammatical functions Prefix: un-, anti-, etc Suffix: -ity, -ation, etc Infix: are inserted inside the stem

Tagalog: um + hingi humingi Circumfixes – precede and follow the stem

English doesn’t stack more affixes. But Turkish can have words with a lot of suffixes. Languages, such as Turkish, tend to string affixes together are

called agglutinative languages.


Surface and Lexical Forms

The surface level of a word represents the actual spelling of that word. geliyorum eats cats kitabım

The lexical level of a word represents a simple concatenation of morphemes making up that word. gel +PROG +1SG eat +AOR cat +PLU kitap +P1SG

Morphological processors try to find correspondences between lexical and surface forms of words. Morphological recognition/ analysis – surface to lexical Morphological generation/ synthesis – lexical to surface


Morphology: Morphemes & Order

Handles what is an isolated form in written text

Grouping of phonemes into morphemes sequence deliverables deliver, able and s (3

units)

Morpheme Combination certain combinations/sequencing possible, other not:

deliver+able+s, but not able+derive+s; noun+s, but not noun+ing typically fixed (in any given language)


Inflectional & Derivational Morphology We can also divide morphology up into two broad classes

Inflectional Derivational

Inflectional morphology concerns the combination of stems and affixes where the resulting word Has the same word class as the original Serves a grammatical/semantic purpose different from the original

After a combination with an inflectional morpheme, the meaning and class of the actual stem usually do not change. eat / eats pencil / pencils

After a combination with an derivational morpheme, the meaning and the class of the actual stem usually change. compute / computer do / undo friend / friendly Uygar / uygarlaş kapı / kapıcı

The irregular changes may happen with derivational affixes.


Morphological Parsing Morphological parsing is to find the lexical form of a word

from its surface form. cats -- cat +N +PLU cat -- cat +N +SG goose -- goose +N +SG or goose +V geese -- goose +N +PLU gooses -- goose +V +3SG catch -- catch +V caught -- catch +V +PAST or catch +V +PP

There can be more than one lexical level representation for a given word. (ambiguity)


Morphological Analysis

Analyzing words into their linguistic components (morphemes). Morphemes are the smallest meaningful units of language.

cars car+PLUgiving give+PROGAsachhilAma AsA+PROG+PAST+1st I/We was/were coming

Ambiguity: More than one alternativesflies flyVERB+PROG

flyNOUN+PLU

mAtAlakare


Fly + s flys flies (y i rule) Duckling

Go-getter get + erDoer do + erBeer ?

What knowledge do we need?How do we represent it?How do we compute with it?


Knowledge needed

Knowledge of stems or roots Duck is a possible root, not ducklWe need a dictionary (lexicon)

Only some endings go on some words Do + er ok Be + er – not ok

In addition, spelling change rules that adjust the surface form Get + er – double the t getter Fox + s – insert e – foxes Fly + s – insert e – flys – y to i – flies Chase + ed – drop e - chased


Put all this in a big dictionary (lexicon)

Turkish – approx 600 106 forms Finnish – 107

Hindi, Bengali, Telugu, Tamil? Besides, always novel forms can be constructed

Anti-missile Anti-anti-missile

Anti-anti-anti-missile ……..

Compounding of words – Sanskrit, German


Morphology: From Morphemes to Lemmas & Categories Lemma: lexical unit, “pointer” to lexicon

typically is represented as the “base form”, or “dictionary headword”

possibly indexed when ambiguous/polysemous: state1 (verb), state2 (state-of-the-art), state3 (government)

from one or more morphemes (“root”, “stem”, “root+derivation”, ...)

Categories: non-lexical small number of possible values (< 100, often < 5-10)


Morphology Level: The Mapping

Formally: A+ 2(L,C1,C2,...,Cn)

A is the alphabet of phonemes (A+ denotes any non-empty sequence of phonemes)

L is the set of possible lemmas, uniquely identified Ci are morphological categories, such as:

grammatical number, gender, case person, tense, negation, degree of comparison, voice, aspect, ... tone, politeness, ... part of speech (not quite morphological category, but...)

A, L and Ci are obviously language-dependent


Morphological Analysis (cont.)

Relatively simple for English. But for many Indian languages, it may be more difficult.

Examples

Inflectional and Derivational Morphology. Common tools: Finite-state transducers


Simple Rules


Adding in the Words


Derivational Rules


Parsing/Generation vs. Recognition

Recognition is usually not quite what we need. Usually if we find some string in the language we need

to find the structure in it (parsing) Or we have some structure and we want to produce a

surface form (production/generation) Example

From “cats” to “cat +N +PL” and back


Finite State Transducers

The simple story Add another tape Add extra symbols to the transitions

On one tape we read “cats”, on the other we write “cat +N +PL”, or the other way around.


FSTs


Transitions

c:c means read a c on one tape and write a c on the other +N:ε means read a +N symbol on one tape and write nothing on the

other +PL:s means read +PL and write an s

c:c a:a t:t +N:ε +PL:s


Typical Uses

Typically, we’ll read from one tape using the first symbol on the machine transitions (just as in a simple FSA).

And we’ll write to the second tape using the other symbols on the transitions.


Ambiguity

Recall that in non-deterministic recognition multiple paths through a machine may lead to an accept state. Didn’t matter which path was actually traversed

In FSTs the path to an accept state does matter since differ paths represent different parses and different outputs will result


Ambiguity

What’s the right parse for Unionizable Union-ize-able Un-ion-ize-able

Each represents a valid path through the derivational morphology machine.


Ambiguity

There are a number of ways to deal with this problem Simply take the first output found Find all the possible outputs (all paths) and return

them all (without choosing) Bias the search so that only one or a few likely paths

are explored


The Gory Details

Of course, its not as easy as “cat +N +PL” <-> “cats”

As we saw earlier there are geese, mice and oxen But there are also a whole host of spelling/pronunciation

changes that go along with inflectional changes Cats vs Dogs Fox and Foxes


Multi-Tape Machines

To deal with this we can simply add more tapes and use the output of one tape machine as the input to the next

So to handle irregular spelling changes we’ll add intermediate tapes with intermediate symbols


Generativity

Nothing really privileged about the directions. We can write from one and read from the other or vice-

versa. One way is generation, the other way is analysis


Multi-Level Tape Machines

We use one machine to transduce between the lexical and the intermediate level, and another to handle the spelling changes to the surface tape


Lexical to Intermediate Level


Intermediate to Surface

The add an “e” rule as in fox^s# <-> foxes#


Foxes


Note

A key feature of this machine is that it doesn’t do anything to inputs to which it doesn’t apply.

Meaning that they are written out unchanged to the output tape.

Turns out the multiple tapes aren’t really needed; they can be compiled away.


Overall Scheme

We now have one FST that has explicit information about the lexicon (actual words, their spelling, facts about word classes and regularity). Lexical level to intermediate forms

We have a larger set of machines that capture orthographic/spelling rules. Intermediate forms to surface forms


Overall Scheme

cs60057 speech &natural language processing

Documents

processing levels knowledge

ambiguous levels

different levels

wordsense ambiguities

wordsense disambiguation

pipeline processing

required linguistic

example of syntactic