Semantics Ling 571 Fei Xia Week 6: 11/1-11/3/05. Outline Meaning representation: what formal structures should be used to represent the meaning of a sentence?

Semantics

Ling 571

Fei Xia

Week 6: 11/1-11/3/05

Outline

• Meaning representation: what formal structures should be used to represent the meaning of a sentence?

• Semantic analysis: how to form the formal structures from smaller pieces?

• Lexical semantics:

Meaning representation

Meaning representation

• Requirements that meaning representations should fulfill

• Types of meaning representation:– First order predicate calculus (FOPC)– Frame-based representation– Semantic network– Conceptual dependency diagram

Requirements

• Verifiability

• Unambiguous representations

• Canonical form

• Inference

• Expressiveness

Verifiability

• A system's ability to compare the state of affairs described by a representation to the state of affairs in some world as modeled in a knowledge base

• Example: – Sent: Maharani serves vegetarian dishes.– Question: Is the statement true?

Unambiguous representation

• Representations should have a single unambiguous interpretation.

• Example:– Mary and John bought a book– Two students met three teachers– A German teacher– A Chinese restaurant– A Canadian restaurant

Canonical form

• Sentences with the same thing should have the same meaning representation

• Example:– Alternations: active/passive, dative shift– Does Maharani have vegetarian dishes?– Do they serve vegetarian food at Maharani?

Inference

• a system's ability to draw valid conclusions based on the meaning representation of inputs and its store of background knowledge.

• Example:– Sent: Maharani serves vegetarian dishes– Question: can vegetarians eat at Maharani?

Expressiveness

• A system should be expressive enough to handle an extremely wide range of subject matter.

• Example: – Belief: I think that he is smart.– Hypothetical statement: If I were you, I would buy that

book.– Former president, fake ID, allegedly, apprarently

Meaning representation

• Requirements – Verifiability– Unambiguous representations– Canonical form– Inference– Expressiveness

• Types of meaning representation:– First order predicate calculus (FOPC)– Frame-based representation– Semantic network– Conceptual dependency diagram

FOPC

• Elements of FOPC

• Representing– Categories– Events– Time (including tense)– Aspect– Belief– …

Elements of FOPC

• Terms:– Constant: specific objects in the world: e.g., Maharani– Variable: a particular unknown object or an arbitrary

object: e.g., a restaurant– Function: concepts: e.g., LocationOf(Maharani)

• Predicates: referring to relations that hold among objects:– Ex: Serve(Maharani, food)– Arguments of predicates must be terms.

Elements of FOPC (cont)

• Logical connectives:

• Quantifier:

• Example: All restaurants serve food.

,,

,

),()(Re foodxServexstaurantx

Inference rules• Modus ponens:

• Conjunction:

• Disjunction:

• Simplification:

• ….

FOPC

• Elements of FOPC

• Representing– Categories– Events– Time– Aspect– Belief– …

Representing time

• Past perfect: I had arrived in NY

• Simple past: I arrived in NY

• Present perfect: I have arrived in NY

• Present: I arrive in NY

• Simple future: I will arrive in NY

• Future perfect: I will have arrived in NY

Representing time (cont)

• Reichenbach’s approach– E: the time of the event– U: the time of the utterance– R: the reference point

• Example:– Past perfect: I had arrived: E > R > U– Simple past: I arrived: E=R > U– Present perfect: I have arrived: E > R=U

Aspect

• Four types of event expression:– Stative: I like books. I have a ticket– Activity: She drove a Mazda. I live in NY– Accomplishment: Sally booked her flight.– Achievement: He reached NY.

• Differences:– Being in a state or not– occurring at a given time, or over some span of a time – Resulting in a state: happening in an instant or not.

Distinguishing four types

• Allowing progressive, imperative– *I am liking books. – *Like books.

• Modified by in-phrase, for-phrase: in a month, for a mont– He lived in NY for five years.– *He reached NY for five minutes.

Distinguishing four types (cont)

• “Stop” test: stop doing something– *He stopped reaching NY.– He stopped booking the ticket

• Modified by adverbs such as “deliberately”, “carefully”– *He likes books deliberately

Representing beliefs

• John believes that Mary ate lunch.• One possibility:

• Another possibility:

),(),(),(Pr

),(),(),(,,

lunchvEatenMaryvEatervuopBelieved

JohnuBelieverEatingvIsAbelievinguIsAvu

)),(,( lunchMaryEatingJohnBelieving

Representing beliefs (cont)

• Substitution does not work• Example:

– John knows Flight 1045 is delayed– Mary is on Flight 1045– Does John know that Mary’s flight was

delayed?

FOPC is not sufficient.Use modal logic

Summary of meaning representation

• Five requirements:– Verifiability– Unambiguous representations– Canonical form– Inference– Expressiveness

• Four types of representations:– First order predicate calculus (FOPC)– Frame-based representation– Semantic network– Conceptual dependency diagram

Outline

• Meaning representation:

• Semantic analysis: how to form the formal structures from smaller pieces?

• Lexical semantics:

Semantic analysis

Semantic analysis

• Goal: to form the formal structures from smaller pieces

• Three approaches:– Syntax-driven semantic analysis– Semantic grammars– Information extraction: filling templates

Syntax-driven approach

• Parsing then semantic analysis, or parsing with semantic analysis.

• Semantic augmentations to grammars (e.g., CFG or LTAG)– Associate FOPC expression with lexical items– Use

– Use complex-terms

ressionexp

)()))((( APAxxP

• Sentence: AyCaramba serves meat• Goal:

• Augmented rules:

),(),(),( MeateServedAyCarambaeServerservingeeIsA

}{

}{

}.{

)}.(.{

)}.(.{

)},(),(),({

meatmeatN

AyCarambaAyCarambaN

semNNNP

semNPsemVPVPNPS

semNPsemVNPVVP

xeServedyeServerservingeIsAeyxservesV

ressionexp

Quantifiers

• Sentence: A restaurant serves meat• Goal:

• Augmented rules:

),(),(

),()Re,(

MeateServedxeServer

ServingeIsAstaurantxIsAxe

),(

)}('..{'

}{

)}.,('

}{

restaurantxIsAxrestauranta

xsemNxsemDetNDetNP

restaurantrestaurantN

semNxIsAxNN

aDet

Complex terms

• Current formula:

• Goal:

• What is needed:

),(),(

),()Re,(

MeateServedxeServer

ServingeIsAstaurantxIsAxe

),())Re,(,(

),(

MeateServedstaurantxxIsAeServer

ServingeIsAe

,...))(...,(,...))(...,(

,...))(...,(,...))(...,(

xPbodyxbodyxP

xPbodyxbodyxP

Quantifier scoping

• Sentence: Every restaurant has a menu• Formula with complex terms

• Reading 1:

• Reading 2:

)),(,(

))Re,(,(),(

menuyIsAyeHad

staurantxIsAxeHaverHavingeIsAe

),(),(),(),(

)Re,(

yeHadxeHaverHavingeIsAmenuyIsAye

staurantxIsAx

)),(),(),((

)Re,(),(

yeHadxeHaverHavingeIsAe

staurantxIsAxmenuyIsAy

Semantic analysis

• Goal: to form the formal structures from smaller pieces

• Three approaches:– Syntax-driven semantic analysis– Semantic grammar– Information extraction: filling templates

Semantic grammar

• Syntactic parse trees only contain parts that are unimportant in semantic processing.

• Ex: Mary wants to go to eat some Italian food

• Rules in a semantic grammar– InfoRequest USER want to go to eat FOODTYPE– FOODTYPENATIONALITY FOODTYPE– NATIONALITYItalian/Mexican/….

Semantic grammar (cont)

Pros:

• No need for syntactic parsing

• Focus on relevant info

• Semantic grammar helps to disambiguate

Cons:

• The grammar is domain-specific.

Information extraction

• The desired knowledge can be described by a relatively simple and fixed template.

• Only a small part of the info in the text is relevant for filling the template.

• No full parsing is needed: chunking, NE tagging, pattern matching, …

• IE is a big field: e.g., MUC. KnowItAll

Summary of semantic analysis

• Goal: to form the formal structures from smaller pieces

• Three approaches:– Syntax-driven semantic analysis– Semantic grammar– Information extraction

Outline

• Meaning representation

• Semantic analysis

• Lexical semantics

Lexical semantics

What is lexical semantics?

• Meaning of word: word senses• Relations among words:

• Predicate-argument structures• Thematic roles• Selectional restrictions

• Mapping from conceptual structures to grammatical functions

• Word classes and alternations

Important resources

• Dictionaries

• Ontology and taxonomy

• WordNet

• FrameNet

• PropBank

• Levin’s English verb classes

• ….

Meaning of words

• Lexeme is an entry in the lexicon that includes– Orthographic form– Phonological form– Sense: lexeme’s meaning

Relations among lexemes

• Homonyms: same orth. and phon. forms, but different, unrelated meanings– bank vs. bank

• Homophones: same phon. different orth– read vs. red, to, two, and too.

• Homographs: same orth, different phon.– bass vs. bass

Polysemy

• Word with multiple but related meanings– He served his time in prison– He served as U.N. ambassador– They rarely served lunch after 3pm.

• What’s the difference between polysemy and homonymy:– Homonymy: distinct, unrelated meanings– Polysemy: distinct but related meanings– How to decide: etymology, notion of coincidence

Synonymy

• Different lexemes with the same meaning

• Substitutable in some environment:– How big is that plane?– How large is that plane?

• What influences substitutablity?– Polysemy: big brother vs. large brother– Subtle shade of meaning: first class fare/?price– Colllocational constraints: big/?large mistake– Register: social factors

Hyponymy

• General: hypernym– “vehicle” is a hypernym of “car”

• Specific: hyponym– “car” is a hyponym of “vehicle”.

• Test: X is a car implies that X is a vehicle.

Ontology and taxonomy

• Ontology: – It is a specification of a conceptualization of a knowledge domain– It is a controlled vocabulary that describes objects and the

relations between them in a formal way, and has strict rules about how to specify terms and relationships.

• Taxonomy: – A taxonomy is a hierarchical data structure or a type of

classification schema made up of classes, where a child of a taxonomy node represents a more restricted, smaller, subclass than its parent.

– a particular arrangement of the elements of an ontology into a tree-like class inclusion structure.

WordNet

• Most widely used lexical database for English

• Developed by George Miller etc. at Princeton

• Three databases: Noun, Verb, Adj/Adv

• Each entry in a database: a unique orthographic form + a set of senses

• Synset: a set of synonyms

• http://www.cogsci.princeton.edu/~wn

WordNet (cont)

• Nouns:– Hypernym: meal, lunch– Has-Member: crew, pilot– Has-part: table, leg– Antonym: leader, follower

• Verbs:– Hypernym: travel, fly– Entail: snoresleep– Antonym: increase decrease

• Adj/Adv:– Antonym: heavy light, quickly slowly

Lexical semantics

• Meaning of word: word senses• Relations among words:

• Predicate-argument structures• Thematic roles• Selectional restrictions

• Mapping from conceptual structures to grammatical functions

• Word classes and alternations

Predicate-argument structure

• Predicate-argument:– Verb/adj as predicate– Nouns etc. as arguments– Example: buy(Mary, book)

• Subcategorization frame: – specify number, position, and syntactic category of arguments

(or complements)– Example:

• (NP, NP): I want Italian food• (NP, Inf-VP): I want to save money• (NP, NP, Inf-VP): I want the book to be delivered tomorrow.

Thematic (Semantic) roles

• A set of roles:– Agent: the volitional causer of an event– Force: the non-volitional causer of an event– Patient/Theme: the one most directly affected by an event– Experiencer: the experiencer of an event– Others: Instrument, Source, Goal, Beneficiary, …

• Example: – John broke a glass– John broke an ankle in the game

Selectional restriction

• Mary ate the cake• ?The table ate the cake

• Mary ate Italian food with her friends.• Mary ate somewhere with her friends.

• White house announced that …• The spider assassinated the fly.

FrameNet

• Developed by Fillmore and Baker at UC Berkeley since 1997.

• http://www.icsi.berkeley.edu/~framenet• FrameNet database has two parts:

– Frame database: a list of semantic frames, and relations between them, such as frame inheritance and frame composition.

– Lexical database: each entry (called a lexical unit) is a (lemma, semantic frame) pair.

Semantic frames

• Definition• Frame elements (FEs): conceptual structure

– Core FEs: Communicator, Medium, Message, Topic– Non-Core FEs: time, place, manner

• Inherit from:• Subframes:• Lexical units: • Example sentences:

One frame

• Frame: Communication– Definition: A Communicator conveys a Message to

an Addressee. the Topic and Medium of the communication also may be expressed.

– Core FEs: Addressee, Communicator, Medium, Message, Topic

– Lexical units: communicate, indicate, signal

Another frame

Frame: Statement– Inherit from: Communication

– Definition: This frame contains verbs and nouns that communicate the act of a Speaker to address a Message to some Addressee using language.

– Core FEs: Communicator, Medium, Message, Topic

– Lexical units: admit, affirm, express,….

Project status

• More than 625 semantic frames, 8900 entries in the lexicon.

• Version 1.2 released in June 2005.

• Book: “FrameNet: Theory and Practice” (printed June 2005)

Proposition Bank (PropBank)

• Developed by Palmer and Marcus at UPenn.

• http://www.cis.upenn.edu/~ace

• Annotate the English Penn Treebank with predicate-argument information

• Corpus can be used for automatic labeling of thematic roles

Semantic tags

• Main tags: – Arg0: Agent– Arg1: theme or direct object– Arg2: instrument, indirect object– …

• Secondary tags: – ArgM-DIR: direction– ArgM-LOC: locative– ArgM-NEG: negation– ArgM-DIS: discourse– …

Semantic tags (cont)• Main tags are defined based on each verb.• Example:

– Buy: John bought a book from Mary for 5 dollars– Sell: Mary sold a book to John for 5 dollars– Pay: John paid Mary 5 dollars for a book.

Arg0 Arg1 Arg2 Arg3

Buy buyer thing bought seller price paid

Sell seller thing bought buyer price paid

Pay buyer price paid seller thing bought

Lexical semantics

• Meaning of word: word senses• Relations among words:

• Predicate-argument structures• Thematic roles• Selectional restrictions

• Mapping from conceptual structure to grammatical function

• Word classes and alternations

Mapping between conceptual structure and grammatical function

• Buy: buyer, thing bought, seller, price,….

• Possible syntactic realizations:– (buyer, thing bought): John bought a book– (price, thing bought): $5 can buy two books– (thing bought, seller): The book was bought from

Mary– (buyer, thing bought, seller): John bought a book from

Mary.

– **(buyer, price): John bought $5.

Alternations

• An alternation is a set of different mappings of conceptual roles to grammatical function.

• Example: dative alternation– John gave Mary a book– John gave a book to Mary

• Verb classes: give, donate,

Levin’s verb classes

• Levin (1993):– Verb classes– Alternations– Show the list of alternatives a verb class can take.

• Problems:– Many verbs appear in multiple classes– Verbs in the same classes do not behave exactly the

same: e.g, (meet, visit), (give, donate),….

Summary of lexical semantics (1)

• Meaning of word: word senses

• Relations among words: – Homonyms: bank, bank– Homophones: read. red– Homographs: bass, bass– Polysemy: bank: blood bank, financial bank– Synonyms: big, large– Hypernym/Hyponym: vehicle, car

• Ontology and taxonomy

• WordNet

Summary of lexical semantics (2)

• Predicate-argument structures

• Thematic roles

• Selectional restrictions

• FrameNet

• PropBank

• Mapping from conceptual structures to grammatical functions

• Word classes and alternations

• Levin’s verb classes for English

Summary of lexical semantics (3)

Summary of semantics

• Meaning representation:– Criteria for good representation– First-order predicate calculus (FOPC)

• Semantic analysis: – Syntax-based semantic analysis– Semantic grammar– Information extraction

• Lexical semantics: – WordNet– FrameNet– PropBank– Levin’s verb classes