FrameNet Meets the Semantic Web

Srini Narayanan

Charles Fillmore

Collin Baker

Miriam Petruck

Outline of Presentation

• Semantic Frames and the FrameNet Project • Status of FrameNet Data and Software • Details on the FrameNet process• Comparison to other ontologies/resources• Afternoon session: Going through the

annotation process demo.

The FrameNet Project

• Phase I (NSF, 1997-2000)– ICSI, U-Colorado– Conceptual basis, used existing tools, and perl

• Phase II (NSF, 2000-2003)– ICSI, U-Colorado, SRI, SDSU– Scaling up, uses SQL database and Java-based

in house tools. Pilot applications developed.

The FrameNet Project

C Fillmore PI (ICSI)Co-PI’s:

S Narayanan (ICSI, SRI)D Jurafsky (U Colorado) J M Gawron (San Diego State U)

Staff: C Baker Project Manager B Cronin Programmer C Wooters Database Designer

Applications

An important goal of our work is to present information about the words in a form that will prove usable in various NLP applications:

1. Question Answering (Berkeley, Colorado)

2. Semantic Extraction (Berkeley, SRI, Colorado)

3. Machine Translation (San Diego State)

Frames and Understanding

• Hypothesis: People understand things by performing mental operations on what they already know. Such knowledge is describable in terms of information packets called frames.

FrameNet in the Larger Context

• The long-term goal is to reason about the world in a way that humans understand and agree with.

• Such a system requires a knowledge representation that includes the level of frames.

• FrameNet can provide such knowledge for a number of domains.

• FrameNet representations complement ontologies and lexicons.

The core work of FrameNet

1. characterize frames

2. find words that fit the frames

3. develop descriptive terminology

4. extract sample sentences

5. annotate selected examples

6. derive "valence" descriptions

Lexicon Building

• We study words,• describe the frames or conceptual structures

which underlie them,• examine sentences that contain them

(from a vast corpus of written English),• and record the ways in which information

from the associated frames are expressed in these sentences.

The Core Data

The basic data on which FrameNet descriptions are based take the form of a collection of annotated sentences, each coded for the combinatorial properties of one word in it. The annotation is done manually, but several steps are computer-assisted.

The Process

• Sentences containing a given word are extracted from the corpus and made available for annotation.

• Student annotators select the phrases that identify particular semantic roles in the sentences, and tag them with the name of these roles.

• Automatic processes then provide grammatical information about the tagged phrases.

SAMPLE ANNOTATIONS

Types of Words / Frames

o events

o artifacts, built objects

o natural kinds, parts and aggregates

o terrain features

o institutions, belief systems, practices

o space, time, location, motion

o etc.

Event Frames

Event frames have temporal structure, and generally have constraints on what precedes them, what happens during them, and what state the world is in once the event has been completed.

Sample Event Frame:Commercial Transaction

Initial state:Vendor has Goods, wants MoneyCustomer wants Goods, has Money

Transition:Vendor transmits Goods to CustomerCustomer transmits Money to Vendor

Final state:Vendor has Money

Customer has Goods

Sample Event Frame:Commercial Transaction

Initial state:Vendor has Goods, wants MoneyCustomer wants Goods, has Money

Transition:Vendor transmits Goods to CustomerCustomer transmits Money to Vendor

Final state:Vendor has Money

Customer has Goods

(It’s a bit more complicated than that.)

Partial Wordlist for Commercial Transactions

Verbs: pay, spend, cost, buy, sell, charge

Nouns: cost, price, payment

Adjectives: expensive, cheap

Meaning and Syntax

The various verbs that evoke this frame introduce the elements of the frame in different ways. The identities of the buyer, seller, goods and

money

Information expressed in sentences containing these verbs occurs in different places in the sentence depending on the verb.

Customer Vendor

Goods Money

BUY

from

for

She bought some carrots from the greengrocer for a dollar.

Customer Vendor

Goods Money

PAY

for

to

She paid a dollar to the greengrocer for some carrots.

Customer Vendor

Goods Money

PAY

for

She paid the greengrocer a dollar for the carrots.

Customer Vendor

Goods Money

SPEND

on

She spent a dollar on the carrots.

Customer Vendor

Goods Money

SELL

for

to

The greengrocer sold some carrots to her for a dollar.

Customer Vendor

Goods Money

SELL

for

The greengrocer sold her some carrots for a dollar.

Customer Vendor

Goods Money

CHARGE

for

The greengrocer charged a dollar for a bunch of carrots.

Customer Vendor

Goods Money

CHARGE

for

The greengrocer charged her a dollar for the carrots.

Customer Vendor

Goods Money

COST

A bunch of carrots costs a dollar.

Customer Vendor

Goods Money

COST

A bunch of carrots cost her a dollar.

Customer Vendor

Goodsto do X

Money

COST

IT

It costs a dollar to ride the bus.

Customer Vendor

Goodsto do X

Money

COST

IT

It cost me a dollar to ride the bus.

FrameNet Product

• For every target word,

• describe the frames or conceptual structures which underlie them,

• and annotate example sentences that cover the ways in which information from the associated frames are expressed in these sentences.

FN work: characterizing frames

• One of the things we do is characterize such information packets - beginning with informal descriptions.

• We can begin with Revenge.

The Revenge frame

The Revenge frame involves a situation in which

a) A has done something to harm B and

b) B takes action to harm A in turn

c) B's action is carried out independently of any legal or other institutional setting

FN work: finding words in frame

• We look for words in the language that bring to mind the individual frames.

• We say that the words evoke the frames.

Vocabulary for Revenge

• Nouns: revenge, vengeance, reprisal, retaliation

• Verbs: avenge, retaliate, revenge, get back (at), get even (with), pay back

• Adjectives: vengeful, vindictive

FN work: choosing FE names

• We develop a descriptive vocabulary for the components of each frame, called frame elements (FEs).

• We use FE names in labeling the constituents of sentences exhibiting the frame.

FEs for Revenge

• Frame Definition: Because of some injuryinjury to something or someone important to an avengeravenger, the avengeravenger inflicts a punishmentpunishment on the offenderoffender. The offenderoffender is the person responsible for the injuryinjury. The injured_partyinjured_party may or may not be the same individual as the avengeravenger.

• FE List: avengeravenger, offenderoffender, injuryinjury, injured_partyinjured_party, punishmentpunishment.

FN work: collecting examples

• We extract from our corpus examples of sentences showing the uses of each word in the frame.

Obviously we need to conduct a more regimented search, grouping examples with related structures.

Examples of simple use are swamped by the idiomatic phrase"with a vengeance".

FN work: annotating examples

• We select sentences exhibiting common collocations and showing all major syntactic contexts.

• Using the names assigned to FEs in the frame, we label the constituents of sentences that express these FEs.

FN work: summarizing results

• Automatic processes summarize the results, linking FEs with information about their grammatical realization.

• The output is presented in the form of various reports in the public website, in XML format in the data release.

I avenged my brother.

I avenged his death.

Querying the data: meaning to form

Through various viewers built on the FN database we can, for example, ask how particular FEs get expressed in sentences evoking a given frame.

By what syntactic means is offender realized?

• Sometimes as direct object:we'll pay you back for that

• Sometimes with the preposition onthey'll take vengeance on you

• Sometimes with againstwe'll retaliate against them

• Sometimes with withshe got even with me

• Sometimes with atthey got back at you

By what syntactic means is offender realized?

• Sometimes as direct object:we'll pay you back for that

• Sometimes with the preposition onthey'll take vengeance on you

• Sometimes with againstwe'll retaliate against them

• Sometimes with withshe got even with me

• Sometimes with atthey got back at you

It's these word-by-wordspecializations inFE-marking that makeautomatic FE recognitiondifficult.

Querying the data: form to meaning

Or, going from the grammar to the meaning, we can choose particular grammatical contexts and ask which FEs get expressed in them.

What FE is expressed by the object of avenge?

• Sometimes it's the injured_partyI've got to avenge my brother

• .Sometimes it's the injuryMy life goal is to avenge my brother's

murder.

Evaluation

• Lexical coverage. We want to get all of the important words associated with each frame.

• Combinatorics. We want to get all of the syntactic patterns in which each word functions to express the frame.

Evaluation

• We do not ourselves collect frequency data. That will wait until methods of automatic tagging get perfected.

• In any case, the results will differ according to the type of corpus - financial news, children's literature, technical manuals, etc.

What do we end up with?

• Frames

• Lexical entries

• Annotations

Sample from frames list

Creating, Crime_scenario, Criminal_investigation, Criminal_process, Cure. Custom, Damaging, Dead_or_alive, Death, Deciding, Deny_permission, Departing, Desirability, Desiring, Destroying, Detaining, Differentiation, Difficulty, Dimension, Direction, Dispersal, Documents, Domain, Duplication, Duration, Eclipse, Education_teaching,Emanating, Emitting, Emotion_active, Emotion_directed, Emotion_heat, Employing, Employment, Emptying, Encoding, Endangering, Entering_of_plea, Entity, Escaping, Evading. Evaluation, Evidence, Excreting, Execution,

Sample from lexical unit list

• * augmentation.N (Expansion)

• * augur.V (Omen)

• * August.N (Calendric_unit)

• * aunt.N (Kinship)

• * auntie.N (Kinship)

• * austere.A (Frugality)

• * austerity.N (Frugality)

• * author.V (Text_creation)

• * authoritarian.A (Strictness)

• * authorization.N (Documents)

• * autobahn.N (Roadways) • * autobiography.N (Text) • * automobile.N (Vehicle)• * autumn.N (Calendric_unit)• * avalanche.N (Quantity)• * avenge.V (Revenge)• * avenger.N (Revenge) • * avenue.N (Roadways) • * aver.V (Statement)

Added Value: frame relatedness

• We have ways of linking frames to each other, through relations of– inheritance– subframe– "using"

• We would like to explore how our frame relationships can be mapped onto ontological relations.

Frame-to-frame relations

• Revenge inherits Punishment/Reward

• Revenge uses the Hostile_encounter frame

• (see existing tentative frame hierarchy)

Added Value: semantic types

• We also have the means of adding semantic types to words, frames and frame elements.

• Some of these:– negative vs. positive

(disaster vs. bonanza),

– punctual vs. stative (arrive vs. reside),

– artifact vs. natural kind (building vs. tree).

Added Value: semantic types

• For the kinds of nouns that occupy particular FE slots in given frames, we should be able to use the WordNet noun taxonomies.

• This is done in some related work

Added Value: support verbs

• In the case of the event nouns, we keep track of which verbs can combine with which nouns to signal occurrences of the frame evoked by the noun. – take a bath (bathe)

– have an argument (argue)

– wreak vengeance,

– take revenge,

– exact retribution.

Can annotation be automated?

Gildea, D & D Jurafsky, 2000, Automatic labeling of semantic roles, Association for Computational Linguistics, Hong Kong.

Mohit & Narayanan, 2003, Semantic Extraction using Wide-coverage lexical resources, HLT-NAACL 2003.

The Database

The information collected from the data (and a certain amount of information inserted manually by the lexicographers) is stored in a MySQL database.

Current Status

• Current: 7700 Lexical Units– FN1: 1600 Lexical units– FN2: 4400 Lexical Units– Created (not yet annotated): 1280 LU– Other : in process, problems, etc.

Current Status

• 500 Frames

• 7700 Lexical Units

• 130,000 Annotated sentences

Data Distribution

Distributed as XML files with accompanying DTDsSeparate files and DTDs for

– Frame and FE data– –Annotation data– Frame relation data

• Easy to parse with standard XML tools. – Approximately 100 research groups have been

authorized to download release 1.0 of the FN data (Oct., 2002).

• Next release scheduled for August, 2003

FrameNet Software Distribution

• All software is pure Java, and can be run on any platform for which a JVM is available

• Has been successfully run on Solaris, Linux, Mac OS X, and Windows 9x/2000 with very minor modifications

• Server and clients currently being used in Barcelona for annotation in Spanish FN.

• We will streamline the installation process if demand warrants

• We plan to publicly release the full software suite in August, 2003.

Multi-Lingual FrameNets

• Spanish FrameNet– Prof. Carlos Subirats, U A Barcelona– Parallel to English FrameNet, using same frames

• German FrameNet– Prof. Manfred Pinkal, U Saarlandes– Complete annotation of existing parsed corpus,– using English frames where possible

• Japanese FrameNet– Prof. Kyoko Ohara, Keio U– Collecting own corpus, building search tools

Some Comparisons

Is FN an ontology?

• Not exactly, but some users use FN frames as an ontology of event types.

Is FN a thesaurus?

Yes, because it groups words into meaning categories, by way of shared membership in frames.

How is FN different from WN?

FN does not explicitly display semantic relations between words of the sort found in WordNet. (synonymy, antonymy, hyponymy, meronymy, etc.)

Furthermore, FN includes many opposing pairs (hot, cold; tall, short) in the same frame.

Are FN annotations a treebank?

• FrameNet accumulates annotations, but FN annotations are mainly sentences in which only one word is analyzed thoroughly.

• Unlike existing treebanks, e.g., U Penn's PropBank, FN has a richer semantics.

Comparison with Dictionaries

American Heritage Dictionary

• avenge v.1. To inflict a punishment or penalty in return for; revenge2. To take vengeance on behalf of

• revenge v.1. To inflict punishment in return for (injury or insult)2. To seek or take vengeance for (oneself or another person); avenge

American Heritage Dictionary

• avenge v.1. To inflict a punishment or penalty in return for; revenge2. To take vengeance on behalf of

• revenge v.1. To inflict punishment in return for (injury or insult)2. To seek or take vengeance for (oneself or another person); avengeThe FEs of the direct objects are expressed prepositionally;

"in return for" marks the injuryinjury; "for" or "on behalf of" marksthe injured_partyinjured_party.

American Heritage Dictionary

• avenge v.1. To inflict a punishment or penalty in return for [ ]; revenge2. To take vengeance on behalf of [ ]

• revenge v.1. To inflict punishment in return for (injury or insult)2. To seek or take vengeance for (oneself or another person); avengerevenge definer added qualifications on the missing

argument, avenge definer didn't.

American Heritage Dictionary

• avenge v.1. To inflict a punishment or penalty in return for; revenge2. To take vengeance on behalf of

• revenge v.1. To inflict punishment in return for (injury or insult)2. To seek or take vengeance for (oneself or another person); avengeavenge definer claims avenge and revenge are

synonym in sense 1; the revenge definer claims avenge and revenge are synonyms in sense 2.

American Heritage Dictionary

• avenge v.1. To inflict a punishment or penalty in return for; revenge2. To take vengeance on behalf of

• revenge v.1. To inflict punishment in return for (injury or insult)2. To seek or take vengeance for (oneself or another person); avenge

revenge definer included "seek vengeance", not supportedby FN examples.

American Heritage Dictionary

• avenge v.1. To inflict a punishment or penalty in return for; revenge2. To take vengeance on behalf of

• revenge v.1. To inflict punishment in return for (injury or insult)2. To seek or take vengeance for (oneself or another person); avenge

Both definers include "take vengeance" in their definitions, asif that's more transparent than the simple verb.

Comparison with WordNet

We make fewer distinctions.

1. revenge, avenge, retaliate -- (take revenge for a perceived wrong; "He wants to avenge the murder of his brother")

2. retaliate, strike back -- (make a counterattack and return like for like, esp. evil for evil; "The Empire strikes back"; "The Giants struck back and won the opener"; "The Israeli army retaliated for the Hamas bombing")

We make fewer distinctions.

1. revenge, avenge, retaliate -- (take revenge for a perceived wrong; "He wants to avenge the murder of his brother")

2. retaliate, strike back -- (make a counterattack and return like for like, esp. evil for evil; "The Empire strikes back"; "The Giants struck back and won the opener"; "The Israeli army retaliated for the Hamas bombing")

Hard to figure out what motivates distinguishing two senses;personal vs. institutional?

We make fewer distinctions.

1. revenge, avenge, retaliate -- (take revenge for a perceived wrong; "He wants to avenge the murder of his brother")

2. retaliate, strike back -- (make a counterattack and return like for like, esp. evil for evil; "The Empire strikes back"; "The Giants struck back and won the opener"; "The Israeli army retaliated for the Hamas bombing")

Like FrameNet, these entries include Definitions and Examples.FrameNet limits its examples to attested sentences from a Corpus.

FN has more detailed syntax.

revenge, avenge, retaliate -- (take revenge for a perceived wrong; "He wants to avenge the murder of his brother")

*> Somebody ----s something

retaliate, strike back -- (make a counterattack and return like for like, esp. evil for evil; "The Israeli army retaliated for the Hamas bombing")

*> Somebody ----s

*> Somebody ----s PPThe WN sentence templates are impoverished structurally and do not indicate the semantic roles. In fact, retaliate is wrongly described as taking a simple object.

FN has more detailed syntax.

revenge, avenge, retaliate -- (take revenge for a perceived wrong; "He wants to avenge the murder of his brother")

*> Somebody ----s something

retaliate, strike back -- (make a counterattack and return like for like, esp. evil for evil; "The Israeli army retaliated for the Hamas bombing")

*> Somebody ----s

*> Somebody ----s PPThe identity of the P in PP is important: strike back atmarks the offenderoffender, as does retaliate against; retaliatefor marks the injuryinjury.

FN has more detailed syntax.

revenge, avenge, retaliate -- (take revenge for a perceived wrong; "He wants to avenge the murder of his brother")

*> Somebody ----s something

retaliate, strike back -- (make a counterattack and return like for like, esp. evil for evil; "The Israeli army retaliated for the Hamas bombing")

*> Somebody ----s

*> Somebody ----s PPWhere WordNet merely shows that the words in thesecond synset can occur intransitively, FN would say something about the anaphoric nature of the omitted offender.

Comparison with ontologies

Switching frames

• Revenge is a simple frame, but neither SUMO nor OpenCYC seem to have any conceptual link to it.

• A particular family of frames that we have concentrated on are those that make up the steps and institutions of Criminal_process.

Complex Frames

• With Criminal_process we have, for example,– sub-frame relations (one frame is a component

of a larger more abstract frame) and – temporal relations (one process precedes

another)

Inferencing

• These are the frames with which we are trying to set up inferencing rules for texts about crime reports. (Details in the presentation later.)

In SUMO

• SUMO (Adam Pease) deals with only the upper ontology, and moves toward our frame along this path, stopping at legal action.– entity

– process

– intentional process

– social interaction

– contest

– legal action

In OpenCYC: ArrestingSomeone

ArrestingSomeone: "A specialization of Social Occurrence and CapturingAnimal. In each instances of ArrestingSomeone a law enforcement officer arrests another person, who is then taken into custody. See the related constant #$HeldCaptive."

Trial

comment : [[Def]] "The subcollection of #$LegalConflict events whose instances are heard and decided by a court and are officiated by a #$Judge."

requiredActorSlots : [[Mon]] plaintiffs [[Mon]] defendants

Legal activities

comment : [[Def]] "The collection of all events performed with the purpose of enforcing laws, that are performed by people officially charged with this this duty. Includes most activities of law enforcement officials (such as police) including detection of crime, identification of offenders, and arrests."

LawEnforcementOfficercomment : [[Def]] "An instance of PersonTypeByOccupation, and a specialization of PersonWithOccupation. Each instance of LawEnforcementOfficer is a person whose job is to detect, stop, and/or punish people engaged in illegal activities. The collection LawEnforcementOfficer includes members of local, state, and special police (e.g., transit police) forces, as well as federal agents (e.g., members of border patrols, national security agents). Consequently, a given instance of Law EnforcementOfficer typically also belongs to one of the following collections: #$StateEmployee, #$LocalGovernment Employee, or NationalGovernmentEmployee (see Public SectorEmployee)."

FrameNet for Applications

• Semantic Web (http://www.semanticweb.org)– FN database in DAML+OIL (

http://www.ai.sri.com/~narayana/frame-desc.daml)

• Semantic Extraction using FrameNet• Frame Simulation and Inference

– Translation from frame structure to a simulation based inference tool (KarmaSIM)

• (COLING 2002)

Talk Outline

• FrameNet

• A DAML + OIL Representation of FrameNet

• An Example: Encoding the Criminal Process Frame

• Web Applications of FrameNet.

• Summary and Future Work

Semantic Web• The World Wide Web (WWW) contains a large

and expanding information base.• HTML is accessible to humans but does not

formally describe data in a machine interpretable form.

• XML remedies this by allowing for the use of tags to describe data (ex. disambiguating crawl)

• Ontologies are useful to describe objects and their inter-relationships.

• DAML+OIL (http://www.daml.org) is an markup language based on XML and RDF that is grounded in description logic and is designed to allow for ontology development, transfer, and use on the web.

FrameNet Entities and Relations

• Frames – Background– Lexical

• Frame Elements (Roles)• Binding Constraints

– Identify• ISA(x:Frame, y:Frame)• SubframeOf (x:Frame, y:Frame)• Subframe Ordering

– precedes• Annotation

A DAML+OIL Frame Class

<daml:Class rdf:ID="Frame"> <rdfs:comment> The most general class </rdfs:comment> <daml:unionOf rdf:parseType="daml:collection"> <daml:Class rdf:about="#BackgroundFrame"/> <daml:Class rdf:about="#LexicalFrame"/> </daml:unionOf></daml:Class>

<daml:ObjectProperty rdf:ID="Name"> <rdfs:domain rdf:resource="#Frame"/> <rdfs:range rdf:resource="&rdf-schema;#Literal"/></daml:ObjectProperty>

DAML+OIL Frame Element

<daml:ObjectProperty rdf:ID= "role"> <rdfs:domain rdf:resource="#Frame"/> <rdfs:range rdf:resource="&daml;#Thing"/></daml:ObjectProperty>

</daml:ObjectProperty> <daml:ObjectProperty rdf:ID="frameElement"> <daml:samePropertyAs rdf:resource="#role"/></daml:ObjectProperty>

<daml:ObjectProperty rdf:ID="FE"> <daml:samePropertyAs rdf:resource="#role"/></daml:ObjectProperty>

FE Binding Relation

<daml:ObjectProperty rdf:ID="bindingRelation"> <rdf:comment> See http://www.daml.org/services </rdf:comment> <rdfs:domain rdf:resource="#Role"/> <rdfs:range rdf:resource="#Role"/></daml:ObjectProperty>

<daml:ObjectProperty rdf:ID="identify"> <rdfs:subPropertyOf rdf:resource="#bindingRelation"/> <rdfs:domain rdf:resource="#Role"/> <daml-s:sameValuesAs rdf:resource="#rdfs:range"/></daml:ObjectProperty>

Subframes and Ordering

<daml:ObjectProperty rdf:ID="subFrameOf"> <rdfs:domain rdf:resource="#Frame"/> <rdfs:range rdf:resource="#Frame"/></daml:ObjectProperty>

<daml:ObjectProperty rdf:ID="precedes"> <rdfs:domain rdf:resource="#Frame"/> <rdfs:range rdf:resource="#Frame"/></daml:ObjectProperty>

Talk Outline

• FrameNet

• A DAML + OIL Representation of FrameNet

• An Example: Encoding the Criminal Process Frame

• Applications of FrameNet.

• Summary and Future Work

The Criminal Process FrameFrame Element Description

Court The court where the process takes place

Defendant The charged individual

Judge The presiding Judge

Prosecution FE indentifies the attorneys’ prosecuting the defendant

Defense Attorneys’ defending the defendant

The Criminal Process Frame in DAML+OIL

<daml:Class rdf:ID="CriminalProcess"> <daml:subClassOf rdf:resource="#BackgroundFrame"/></daml:Class>

<daml:Class rdf:ID="CP"> <daml:sameClassAs rdf:resource="#CriminalProcess"/></daml:Class>

DAML+OIL Representation of the Criminal Process Frame Elements

<daml:ObjectProperty rdf:ID="court"> <daml:subPropertyOf rdf:resource="#FE"/> <daml:domain rdf:resource="#CriminalProcess"/> <daml:range rdf:resource="&CYC;#Court-Judicial"/></daml:ObjectProperty>

<daml:ObjectProperty rdf:ID="defense"> <daml:subPropertyOf rdf:resource="#FE"/> <daml:domain rdf:resource="#CriminalProcess"/> <daml:range rdf:resource="&SRI-IE;#Lawyer"/></daml:ObjectProperty>

FE Binding Constraints

<daml:ObjectProperty rdf:ID="prosecutionConstraint"> <daml:subPropertyOf rdf:resource="#identify"/> <daml:domain rdf:resource="#CP.prosecution"/> <daml-s:sameValuesAs rdf:resource="#Trial.prosecution"/></daml:ObjectProperty>

• The idenfication contraints can be between • Frames and Subframe FE’s.• Between Subframe FE’s

• DAML does not support the dot notation for paths.

Criminal Process Subframes

<daml:Class rdf:ID="Arrest"> <rdfs:comment> A subframe </rdfs:comment> <rdfs:subClassOf rdf:resource="#LexicalFrame"/></daml:Class>

<daml:Class rdf:ID="Arraignment"> <rdfs:comment> A subframe </rdfs:comment> <rdfs:subClassOf rdf:resource="#LexicalFrame"/></daml:Class>

<daml:ObjectProperty rdf:ID="arraignSubFrame"> <rdfs:subPropertyOf rdf:resource="#subFrameOf"/> <rdfs:domain rdf:resource="#CP"/> <rdfs:range rdf:resource="#Arraignment"/></daml:ObjectProperty>

Specifying Subframe Ordering

<daml:Class rdf:about="#Arrest">

<daml:subClassOf>

<daml:Restriction>

<daml:onProperty rdf:resource="#precedes"/>

<daml:hasClass rdf:resource="#Arraignment"/>

</daml:Restriction>

</daml:subClassOf>

</daml:Class>

DAML+OIL CP Annotations

<fn:Annotation> <tpos> "36352897" </tpos> <frame rdf:about ="&fn;Arrest"> <time> In July last year </time> <authorities> a German border guard </authorities> <target> apprehended </target> <suspect> two Irishmen with Kalashnikov assault rifles. </suspect> </frame></fn:Annotation>

Current Status of DAML Encoding

• All FrameNet 1 data is available in DAML+OIL– annotations

– frame descriptions.

• The translator has also been updated to handle the more complex semantic relations (both frame and frame element based) in FrameNet 2.

• We plan to release both the XML and the DAML+OIL versions of all FrameNet 2 releases.

Talk Outline

• FrameNet

• A DAML + OIL Representation of FrameNet

• An Example: Encoding the Criminal Process Frame

• Applications of FrameNet.

• Summary and Future Work

FrameNet for Applications

• Semantic Web (http://www.semanticweb.org)– FN database in DAML+OIL (

http://www.ai.sri.com/~narayana/frame-desc.daml)

• Semantic Extraction using FrameNet• Or can FrameNet be automated

• Frame Simulation and Inference – Translation from frame structure to a simulation based

inference tool (KarmaSIM) • (COLING 2002)

Semantic Extraction

• Behrang Mohit and Srini Narayanan– HLT-NAACL 2003.

Enhancing IE Techniques

• IE techniques currently use no inference (mostly!) – Robert Pickett was charged with felony possession of a

handgun and sentenced to 5 years in a federal prison.• Says Pickett was arrested

• Frame-based inferences can be useful for a variety of applications including individual/topic tracking, bridging inferences/co-reference resolution.

• FrameNet subframe structure and bindings can be exploited for this purpose.

A Simulation Semantics for Inference

• Frame Structure and bindings specify parameters for a simulation/enactment of the event

• Based on previous work (IJCAI 99, AAAI 99, CogSci 2000,

COLING 2002, WWW 2002) – using an “X-schema” based representation, we simulate

the temporal and inferential structure of the Frame-Element and Frame/Subframe relations from FrameNet.

– Direct translation from both the mySQL FN database and the DAML+OIL representation

Reasoning about Events for NL applications (QA, NLU)

• Reasoning about dynamics– Complex event structure

• Multiple stages, interruptions, resources, framing– Evolving events

• Conditional events, presuppositions.– Nested temporal and aspectual references

• Past, future event references– Metaphoric references

• Use of motion domain to describe complex events.• Reasoning with Uncertainty

– Combining Evidence from Multiple, unreliable sources– Non-monotonic inference

• Retracting previous assertions• Conditioning on partial evidence

Previous work• Models of event structure that are able to deal with the

temporal and aspectual structure of events• Models frame-based and metaphoric inference about event

structure.• Based on an active semantics of events and a factorized

graphical model of complex states.– Models event stages, embedding, multi-level perspectives and

coordination.– Event model based on a Stochastic Petri Net representation with

extensions allowing hierarchical decomposition.– State is represented as a Temporal Bayes Net (T(D)BN).– The Event-State representation requires branching time bayes nets

with synchronization or Coordinated Bayes Nets (CBN)

States• Factorized Representation of State uses

Dynamic Belief Nets (DBN’s)– Probabilistic Semantics– Structured Representation

States and Domain Knowledge

• Factorized Representation using Dynamic Belief Nets (DBN’s)– Probabilistic

Semantics

– Structured Representation

Active Event Representations

• Actions and events are coded in active representations called x-schemas which are extensions to Stochastic Petri nets.

• x-schemas are fine-grained and can be used for monitoring and control as well as for inference.

• Badler’s (U Penn) group uses same idea for commanding simulated robots (Jack). Nils Nilsson (SU) uses a similar idea for robot planning called Teleo-Reactive programs.

• Semantic basis for DAML-S, process descriptions of the Semantic Web

Compositional Primitives

process

atomicprocess

compositeprocess

inputs (conditional) outputs preconditions (conditional) effects

controlconstructs

composedBy

whilesequence

If-then-else

fork

...

Sequence: P1;P2

start finish

Done(P1;P2)Atomic

ProcessP2

Done(P1)Atomic

ProcessP1

Ready

Fork: P1|| P2

start finish

Done(P1 || P2)

AtomicProcess

P2

Ready(P1)Atomic

ProcessP1

Ready(P2)

Concurrent-Sync

Done(P2)

Done(P1)

start finish

AtomicProcess

P2

Ready(P1)Atomic

ProcessP1

Ready(P2)

Implementation

DAML-S translation to the modeling environment KarmaSIM [Narayanan, 97] (http://www.icsi.berkeley.edu/~snarayan)

Basic Program:

Input: DAML-S description of Frame relations

Output: Network Description of Frames in KarmaSIM

Procedure:• Recursively construct a sub-network for each control construct.

Bottom out at atomic frame.• Construct a net for each atomic frame• Return network

A Precise Notion of Contingency Relations

Activation:Executing one schema causes the enabling, start or continued execution of another schema. Concurrent and sequential activation.Inhibition:Inhibitory links prevent execution of the inhibited x-schema by activating an inhibitory arc. The model distinguishes between concurrent and sequential inhibition, mutual inhibition and aperiodicity.Modification:The modifying x-schema results in control transition of the modified xschema. The execution of the modifying x-schema could result in theinterruption, termination, resumption of the modified x-schema.

Results of Model• Captures fine grained distinctions needed for

interpretation– Frame-based Inferences (COLING02)– Aspectual Inferences (Cogsci98, CogSci01, IJCAI 99,

CL03)– Metaphoric Inferences (AAAI99)– Biological Evidence (CogSci03, BL03)

• Sufficient Inductive bias for verb learning (Bailey97, CogSci99), construction learning (Chang03, to Appear)

• Model for DAML-S (ISWC02, WWW02, Computer Networks 03)

Distributed OPErational (DOPE) Semantics

Maps Situation Calculus action axiomatization to CBN Formalism [Narayanan 99, NM2002, NM2003]

Features of CBN representation

Can deal with quantitative information & resources

Natural representation of stochastic actions (selection and effects)

Variety of well established analysis and simulation techniques including

mappings to other logics of change.

Natural representation of change, concurrency, and synchronization

Execution semantics

Problems with T(D)BN• Scaling up to relational structures

• Supports linear (sequence) but not branching (concurrency, coordination) dynamics

Structured Probabilistic Inference

Probabilistic inference for Events

– Filtering• P(X_t | o_1…t,X_1…t)• Update the state based on the observation sequence and state set

– MAP Estimation• Argmaxh1…hnP(X_t | o_1…t, X_1…t)• Return the best assignment of values to the hypothesis variables given

the observation and states– Smoothing

• P(X_t-k | o_1…t, X_1…t)• modify assumptions about previous states, given observation

sequence and state set– Projection/Prediction/Reachability

• P(X_t+k | o_1..t, X_1..t)

Open-Source FrameNet• Use the idea of open source Linux development

– Frame hackers around the world– Distributed vanguard and peer review process– Pilot projects in large social networks (ICSI BCIS

project)

• Develop software and infrastructure – Frame Creation and Modification– Annotation structures– Common API for semantic resources.– Specialized domain FrameNets

Summary The FrameNet Project is making good progress toward our

goal of producing a lexicon for a significant number of English words with uniquely detailed information about their argument structure and the semantics associated with it.

We have an automatic translation from FrameNet to computational representations that Are able to translate FN annotations and frame structure for use by

Semantic Web researchers and use ontologies on the web for semantic typing of FE’s.

Translates Frame representations to a simulation semantics that can perform frame-based inference and may provide a scalable semantics for NL systems.

Ongoing Work: Question Answering• As part of the AQUAINT program (UCB, ICSI,

Stanford), we are tasked with – coming up with a uniform formalism to encode frames,

schemas and metaphors (ScaNaLU 2002)– Designing inference algorithms to reason with semantic

schemas. – Others (UCB, Stanford) are tasked with trying to identify

semantic relations from text. – One possible interchange language choice is DAML-S/OWL-

S

• Hypothesis: Simulation based inference over semantic relations is useful for question answering.

http://www.icsi.berkeley.edu/framenet

http://www.icsi.berkeley.edu/NTL

http://www.icsi.berkeley.edu/framenet

http://www.icsi.berkeley.edu/NTL