november 2009hlt - sentence grammar1 human language technology sentence grammar

November 2009 HLT - Sentence Grammar 1

Human Language Technology

Sentence Grammar


Introduction

This lecture has several themes:

1. Crash course in sentence-level grammar• Jurafsky and Martin 2nd ed. Chapter 12

• Internet Grammar of Englishhttp://www.ucl.ac.uk/internet-grammar/

2. Show how different linguistic phenomena can be captured by grammar rules.

3. Dependency Parsing

4. Tagsets and Treebanks


Part 1

Grammar of English


Different Kinds of Rule

• Morphological rules.. govern how words may be composed: re+invest+ing = reinvesting.

• Syntactic rules .. govern how words and constituents combine to form grammatical sentences.

• Semantic rules .. govern how meanings may be combined.


Syntax: Why?

• You need knowledge of syntax in many applications:– Parsing– Grammar checkers– Question answering/database access– Information extraction– Generation– Translation

• Full versus superficial analysis?


Levels of Grammar Organisation• Word Classes: different parts of speech (POS).• Phrase Classes: sequences of words inheriting the

characteristics of certain word classes.• Clause Classes: sequences of phrases containing

at least one verb phrase.On the basis of these one may define:• Grammatical Relations: role played by

constitutents e.g. subject; predicate; object• Syntax-Semantics interface: mapping between

syntactic structures and meaning


Word Classes• Closed classes.

– determiners : the, a, an, four.– pronouns : it, he etc.– prepositions : by, on, with .– conjunctions : and, or, but.

• Open classes.– nouns refer to objects or concepts: cat , beauty , Coke.– adjectives describe or qualify nouns: fried chickens.– verbs describe what the noun does: John jumps.– adverbs describe how it is done: John runs quickly.


Word Class Characteristics• Different word classes have characteristic

subclasses and propertiesSubclasses Properties

Nounproper; mass;

count

number;

gender

Verbtransitive;

intransitive

Number; gender;

person, tense

Adjectivedimension; age;

colour

Number,

gender


Phrases

• Longer phrases may be used rather than a single word, but fulfilling the same role in a sentence.– Noun phrases refer to objects: four fried chickens.– Verb phrases state what the noun phrase does: kicks the

dog.– Adjective phrases describe/qualify an object:

sickly sweet.– Adverbial phrases describe how actions are done:

very carefully.– prepositional phrases: add information to a verb

phrase: on the table


Phrases can be Complexe.g. Noun Phrases

• Proper Name or Pronoun: Monday; it

• Specifier, noun: the day

• Specifiers, premodifier, noun:the first wet day

• Specifiers, premodifier, noun, postmodifier:The first wet day that I enjoyed in June


But they all fit the same context

• Monday• It• The day• The first wet day• The first wet day that I

enjoyed in June

was sunny.


Clauses• A clause is a combination of noun phrases and

verb phrases• Clauses can exist at the top level (main clause) or

can be embedded (subordinate clause)– Top level clause is a sentence. E.g.

The cat ate the mouse.– Embedded clause is subordinate e.g.

John said that Sandy is sick.

• Unlike phrases, whole sentences can be used to say something complete, e.g. to state a fact or ask a question.


Different Kinds of Sentences

• Assertion: John ate the cat.

• Yes/No question: Did John eat the cat?

• Wh- question: What did John eat?

• Command: Eat the cat John!

• NB. All these forms share the same underlying semantic proposition.


Part II

Context Free Grammar Rules


Formal Grammar

• A formal grammar consists of– Terminal Symbols (T)– Non Terminal Symbols (NT, disjoint from TS)– Start Symbol (a distinguished NT)– Rewrite rules of the form , where and

are strings of symbols


Classes of Grammar

Type Grammars Languages Machines

0 Phrase StructureRecursivelyEnumerable

TM

1 Context SensitiveContextSensitive

LBA

2 Context Free Context Free PDA

3 Regular Regular FSA


Classes of Grammar

• Learnability

• Different classes of grammar result from various restrictions on the form of rules


Restrictions on Rules

For all rules • Type 0 (unrestricted): no restrictions

• Type 1 (context sensitive): ||||

• Type 2 (context free): is a single NT symbol

• Type 3 (regular)– Every rule is of the form A aB or A a

where A,B NT and aT


Which Class for NLP?

• Type 3 (Regular). Good for morphology. Cannot handle central embedding of sentences.The man that John saw eating died.

• Type 2 (Context Free). OK but problems handling certain phenomena e.g. agreement.

• Type 1 (Context Sensitive). Computational properties not well understood. Too powerful.

• Type 0 (Turing). Too powerful.


Weak versus Strong

• Grammar class that is too restrictive– cannot characterise/discriminate exactly NL

sentence structures.

• Grammar class that is too general– has the power to characterise/discriminate

structures that don't exist in human languages.

• More general, higher complexity→ less efficient computations.


Example Grammar

Cabinet discusses police chief’s caseFrench gunman kills fours np vpnp nnp adj nnp n npvp v np


Classifying the Symbols• NT – symbols appearing on the left• Start – symbol appearing only on the left from

which every other symbol can be derived.• T – symbols appearing only on the right• To include words we also need special rules

such as n [police]n [gunman]n [four]

• Latter rules define the lexicon or “dictionary interface”.


Grammar InducesPhrase Structure

French gunman kills four

adj n v n

np np

vp s


Phrase Structure

• PS includes information about – precedence between constituents– dominance between constituents

• PS constitutes a trace of the rule applications used to derive a sentence

• PS does not tell you the order in which the rules were used


Procedural versus Declarative

• A grammar induces a structure but does not tell you how to discover that structure

• A grammar is declarative

• A parser is a procedure that, given a suitable representation of a grammar and a sentence, actually discovers the structure(s).

• A parser is procedural


Handling Linguistic Phenomena

• Different sentence-types

• Nested structures

• Agreement

• Multiwords

• Subcategories of verb


Different Sentence Types........Different Grammar Rules

• DeclarativesJohn left.S → NP VP

• ImperativesLeave!S →VP

• Yes-No QuestionsDid John leave?S →Aux NP VP

• WH QuestionsWhen did John leave?S →Wh-word Aux NP VP


Recursively NestedStructures handled by ....

• Flights to Miami• Flights to Miami from Boston• Flights to Miami from Boston in April• Flights to Miami from Boston in April on

Friday• Flights to Miami from Boston in April on

Friday under $300.• Flights to Miami from Boston in April on

Friday under $300 with lunch.


Recursive Rules

NP → N

NP → NP PP

PP → Preposition NP

Flights from miami to boston


Ambiguity

np np pppp prep np

(the man) (on the hill with a telescope by the sea)

(the man on the hill) (with a telescope by the sea)

(the man on the hill with a telescope)( by the sea)

etc.


Handling Agreement

• NP → Determiner N• Include these days, this day• Exclude this days, these day

NP → NPSingNP → NPPlurNPPlur → DetSing NSingNPPlur → DetPlur NPlur

• Agreement also includes number, gender, case.• Danger: proliferation of categories/rules.


Handling Multiwords

• John ran up the stairs

• John rang up the doctor

• John ran the stairs up*

• John rang the doctor up

• John rang the doctor who lives in Paris up


Ordinary CF rules don’t work

John rang up the doctor• VP → V NP• here V is multiwordJohn rang the doctor up• VP → V NP particle_from _V• here, multiword has split into two parts• challenge is to express the relation between the

parts


Subcategorisation

• Intransitive verb: no objectJohn disappearedJohn disappeared the cat*

• Transitive verb: one objectJohn opened the windowJohn opened*

• Ditransitive verb: two objectsJohn gave Mary the bookJohn gave Mary*


Subcategorisation Rules

• Intransitive verb: no objectVP → V

• Transitive verb: one objectVP → V NP

• Ditransitive verb: two objectsVP → V NP NP

• If you take account of the category of items following the verb, there are about 40 different patterns like this in English.


Overgeneration

• A grammar should generate only sentences in the language.

• It should exclude sentences not in the language.

s n vpvp vn [John]v [snore]v [snores]


Undergeneration

• A grammar should generate all sentences in the language.

• There should not be sentences in the language that are not generated by the grammar.

s n vpvp vn [John]n [gold]v [found]


Appropriate Stuctures

– A grammar should assign linguistically plausible structures.

n v d a nJohn ate a juicy hamburger

vp

s

n v d a nJohn ate a juicy hamburger

np vp

s

np

vs.


Criteria for Evaluating Grammars

• Does it undergenerate? • Does it overgenerate?• Does it assign appropriate structures to sentences

it generates?• Is it simple to understand? How many rules are

there?• Does it contain generalisations or is it just a

collection of special cases?• How ambiguous is it?


Tagsets

• The main parts of speech reflect naturally occurring occurrence data.

• Practical applications often make use of special tags which include additional information such as number and case.

• One of the most commonly used tagsets is the 45-tag Penn Treebank tagset, used for the Brown corpus.


Penn Treebank Tagset


POS Tagging

The grand jury commented on a number of other topics

The/DT grand/JJ jury/NN commmented/VBD

on/IN a/DT number/NN of/IN other/JJ topics/NNS ./.

POS Tagger


Treebanks• Treebanks are corpora in which each sentence has been

paired with a parse tree (presumably the right one).• These are generally created

– By first parsing the collection with an automatic parser– And then having human annotators correct each parse

as necessary.• This requires detailed annotation guidelines that provide a

– POS tagset,

– a grammar and

– instructions for how to deal with particular grammatical constructions.


Penn Treebank• Penn Treebank is a widely used treebank

maintained by the Linguistic Data Consortium.• The Penn Treebank Project annotates naturally-

occurring text for linguistic structure.• Contains skeletal parses showing rough syntactic

and semantic information -- a bank of linguistic trees.

• Most well known is the Wall Street Journal section containing 1M words from the 1987-1989 Wall Street Journal.


Penn Treebank Example


Treebank Grammars

• Treebanks implicitly define a grammar for the language covered in the treebank.

• Simply take the local rules that make up the sub-trees in all the trees in the collection and you have a grammar.

• Not complete, but if you have decent size corpus, you’ll have a grammar with decent coverage.


Treebank Grammars

• Such grammars tend to be very flat due to the fact that they tend to avoid recursion.

• For example, the Penn Treebank has 4500 different rules for VPs. Among them...


Heads in Trees

• Finding heads in treebank trees is a task that arises frequently in many applications.– Particularly important in statistical

parsing

• We can visualize this task by annotating the nodes of a parse tree with the heads of each corresponding node.


Lexically Decorated Tree


Head Finding

• The standard way to do head finding is to use a simple set of tree traversal rules specific to each non-terminal in the grammar.


Noun Phrases


Treebank Uses

• Treebanks (and headfinding) are particularly critical to the development of statistical parsers.

• Also valuable for Corpus Linguistics – Investigating the empirical details of various

constructions in a given language


Dependency Grammars• In CFG-style phrase-structure grammars the main

focus is on constituents.• But it turns out you can get a lot done with just

binary relations among the words in an utterance.• In a dependency grammar framework, a parse is a

tree where – the nodes stand for the words in an utterance– The links between the words represent dependency

relations between pairs of words.– Relations may be typed (labeled), or not.


Dependency Relations


Dependency Parse

They hid the letter on the shelf


Dependency Parsing

• The dependency approach has a number of advantages over full phrase-structure parsing.– Deals well with free word order languages

where the constituent structure is quite fluid– Parsing is much faster than CFG-based parsers– Dependency structure often captures the

syntactic relations needed by later applications

• CFG-based approaches often extract this same information from trees anyway.


Dependency Parsing

• There are two modern approaches to dependency parsing– Optimization-based approaches that search a

space of trees for the tree that best matches some criteria

– Shift-reduce approaches that greedily take actions based on the current word and state.


Summary• Context-free grammars can be used to model various facts

about the syntax of a language.• When paired with parsers, such grammars consititute a

critical component in many applications.• Constituency is a key phenomenon easily captured with

CFG rules.– But certain linguistic phenomena pose significant

problems• Dependency parsing is suitable for languages with free

word order• Treebanks pair sentences in corpus with their

corresponding trees.

november 2009hlt - sentence grammar1 human language technology sentence grammar

Documents

hlt sentence grammar3

hlt sentence grammar11

hlt sentence grammar5

gender slide

grammar of english slide

noun phrases proper

phrase classes

adjective phrases