human language technology for the semantic web //gate.ac.uk/ //nlp.shef.ac.uk/ hamish cunningham...

Human Language Technology for the Semantic Web

http://gate.ac.uk/ http://nlp.shef.ac.uk/

Hamish CunninghamKalina BontchevaDiana Maynard Valentin Tablan

ESWS, Crete, May 2004

[This work has been supported by AKT (http://aktors.org/) and

SEKT (http://sekt.semanticweb.org/)]

http://gate.ac.uk/

http://nlp.shef.ac.uk/

2(130)

Are you wasting your time?

3(130)

Structure of the Tutorial

1. Motivation, background

2. Information Extraction - definition

3. Evaluation – corpora & metrics

4. IE approaches – some examples– Rule-based approaches– Learning-based approaches

5. Semantic Tagging– Using “traditional” IE– Ontology-based IE– Platforms for large-scale processing

6. Language Generation

[Slides: http://gate.ac.uk/sale/talks/esws2004-tutorial.ppt]

4(130)

The Knowledge Economy and Human Language

Gartner, December 2002: • taxonomic and hierarchical knowledge mapping and

indexing will be prevalent in almost all information-rich applications

• through 2012 more than 95% of human-to-computer information input will involve textual language

A contradiction: formal knowledge in semantics-based systems vs. ambiguous informal natural language

The challenge: to reconcile these two opposing tendencies

5(130)

HumanLanguage

Formal Knowledge(ontologies andinstance bases)

(A)IE

CLIE

(M)NLG

ControlledLanguage

OBIE

SemanticWeb; Semantic Grid;Semantic Web Services

KEYMNLG: Multilingual Natural Language GenerationOBIE: Ontology-Based Information ExtractionAIE: Adaptive & Mixed-Initiative IECLIE: Controlled Language IE

HLT & Knowledge: Closing the Language Loop

6(130)

Background and Examples (1)• Like other areas of computer science, HLT has

typical data structures and infrastructure requirements

• Annotation: associating arbitrary data with areas of text or speech

• Defacto standard: Stand-off Markup (e.g. TEI/XCES, NITE, ATLAS, GATE)

• Other issues: visualisation and editing; persistence and search; metrics; component model; baseline NLP tools; ...

• To cut a long story short: HLT has a lot of T underneath it which comes in many shapes and sizes

7(130)

Background and Examples (2)Infrastructure & (many) examples in this tutorial:

• GATE, a General Architecture for Text Engineering: architecture, framework & IDE

Why?• I happen to know a little about it • Free software, relatively comprehensive, widely

used, has extensive Semantic Web support• It means we can ignore the infrastructural

issuesNot a claim that it is the best or only in all cases!

9(130)

Information Extraction (1)

• Information Extraction (IE) pulls facts and structured information from the content of large text collections.

• Contrast IE and Information Retrieval• NLP history: from NLU to IE

(if you can’t score, why not move the goalposts?)

10(130)

Information Extraction (2)• When you can measure what you are speaking about,

and express it in numbers, you know something about it; but when you cannot measure it, when you cannot express it in numbers, your knowledge is of a meager and unsatisfactory kind: it may be the beginning of knowledge, but you have scarcely in your thoughts advanced to the stage of science. (Kelvin)

• Not everything that counts can be counted, and not everything that can be counted counts. (Einstein)

• IE progress driven by quantitative measures

• MUC: Message Understanding Conferences • ACE: Automatic Content Extraction

11(130)

MUC-7 tasks

Held in 1997, around 15 participants inc. 2 UK. Broke IE down into component tasks:

• NE: Named Entity recognition and typing

• CO: co-reference resolution • TE: Template Elements • TR: Template Relations • ST: Scenario Templates

12(130)

“The shiny red rocket was fired on Tuesday. It is the

brainchild of Dr. Big Head. Dr. Head is a staff scientist at

We Build Rockets Inc.”

An Example

• ST: rocket launch event with various participants

• NE: "rocket", "Tuesday", "Dr. Head“, "We Build Rockets"

• CO:"it" = rocket; "Dr. Head" = "Dr. Big Head"• TE: the rocket is "shiny red" and Head's

"brainchild". • TR: Dr. Head works for We Build Rockets Inc.

13(130)

Performance levels

• Vary according to text type, domain, scenario, language

• NE: up to 97% (tested in English, Spanish, Japanese, Chinese, etc. etc.)

• CO: 60-70% resolution • TE: 80% • TR: 75-80% • ST: 60% (but: human level may be

only 80%)

14(130)

What are Named Entities?

• NE involves identification of proper names in texts, and classification into a set of predefined categories of interest

• Person names• Organizations (companies, government

organisations, committees, etc)• Locations (cities, countries, rivers, etc)• Date and time expressions

15(130)

What are Named Entities (2)

• Other common types: measures (percent, money, weight etc), email addresses, Web addresses, street addresses, etc.

• Some domain-specific entities: names of drugs, medical conditions, names of ships, bibliographic references etc.

• MUC-7 entity definition guidelines [Chinchor’97]

http://www.itl.nist.gov/iaui/894.02/related_projects/muc/proceedings/ne_task.html



16(130)

What are NOT NEs (MUC-7)

• Artefacts – Wall Street Journal• Common nouns, referring to named entities –

the company, the committee • Names of groups of people and things named

after people – the Tories, the Nobel prize• Adjectives derived from names – Bulgarian,

Chinese• Numbers which are not times, dates,

percentages, and money amounts

17(130)

Basic Problems in NE

• Variation of NEs – e.g. John Smith, Mr Smith, John.

• Ambiguity of NE types: John Smith (company vs. person) – May (person vs. month) – Washington (person vs. location) – 1945 (date vs. time)

• Ambiguity with common words, e.g. "may"

18(130)

More complex problems in NE

• Issues of style, structure, domain, genre etc.

• Punctuation, spelling, spacing, formatting, ... all have an impact:

Dept. of Computing and MathsManchester Metropolitan UniversityManchesterUnited Kingdom

Tell me more about Leonardo

Da Vinci

19(130)









20(130)

Corpora and System Development• “Gold standard” data created by manual annotation• Corpora are divided typically into a training and testing portion • Rules and/or learning algorithms are developed or trained on

the training part• Tuned on the testing portion in order to optimise

– Rule priorities, rules effectiveness, etc.– Parameters of the learning algorithm and the features used

(typical routine: 10-fold cross validation)• Evaluation set – the best system configuration is run on this

data and the system performance is obtained• No further tuning once evaluation set is used!

21(130)

Some NE Annotated Corpora

• MUC-6 and MUC-7 corpora - English• CONLL shared task corpora

http://cnts.uia.ac.be/conll2003/ner/ - NEs in English and German

http://cnts.uia.ac.be/conll2002/ner/ - NEs in Spanish and Dutch

• TIDES surprise language exercise (NEs in Cebuano and Hindi)

• ACE – English - http://www.ldc.upenn.edu/Projects/ACE/

http://cnts.uia.ac.be/conll2003/ner/

http://cnts.uia.ac.be/conll2002/ner/

http://www.ldc.upenn.edu/Projects/ACE/



22(130)

The MUC-7 corpus• 100 documents in SGML • News domain

Named Entities:• 1880 Organizations (46%)• 1324 Locations (32%)• 887 Persons (22%)• Inter-annotator agreement very high (~97%)• http://www.itl.nist.gov/iaui/894.02/related_project

s/muc/proceedings/muc_7_proceedings/marsh_slides.pdf

http://www.itl.nist.gov/iaui/894.02/related_projects/muc/proceedings/muc_7_proceedings/marsh_slides.pdf



23(130)

The MUC-7 Corpus (2)

<ENAMEX TYPE="LOCATION">CAPE CANAVERAL</ENAMEX>, <ENAMEX TYPE="LOCATION">Fla.</ENAMEX> &MD; Working in chilly temperatures <TIMEX TYPE="DATE">Wednesday</TIMEX> <TIMEX TYPE="TIME">night</TIMEX>, <ENAMEX TYPE="ORGANIZATION">NASA</ENAMEX> ground crews readied the space shuttle Endeavour for launch on a Japanese satellite retrieval mission.

<p>Endeavour, with an international crew of six, was set to blast off from

the <ENAMEX TYPE="ORGANIZATION|LOCATION">Kennedy Space Center</ENAMEX> on <TIMEX TYPE="DATE">Thursday</TIMEX> at <TIMEX TYPE="TIME">4:18 a.m. EST</TIMEX>, the start of a 49-minute launching period. The <TIMEX TYPE="DATE">nine day</TIMEX> shuttle flight was to be the 12th launched in darkness.

24(130)

ACE – Towards Semantic Tagging of Entities

• MUC NE tags segments of text whenever that text represents the name of an entity

• In ACE (Automated Content Extraction), these names are viewed as mentions of the underlying entities. The main task is to detect (or infer) the mentions in the text of the entities themselves

• Rolls together the NE and CO tasks• Domain- and genre-independent approaches• ACE corpus contains newswire, broadcast news

(ASR output and cleaned), and newspaper reports (OCR output and cleaned)

25(130)

ACE Entities

• Dealing with – Proper names – e.g., England, Mr. Smith, IBM– Pronouns – e.g., he, she, it– Nominal mentions – the company, the spokesman

• Identify which mentions in the text refer to which entities, e.g., – Tony Blair, Mr. Blair, he, the prime minister, he– Gordon Brown, he, Mr. Brown, the chancellor

26(130)

ACE Example <entity ID="ft-airlines-27-jul-2001-2" GENERIC="FALSE" entity_type = "ORGANIZATION"> <entity_mention ID="M003" TYPE = "NAME" string = "National Air Traffic Services"> </entity_mention> <entity_mention ID="M004" TYPE = "NAME" string = "NATS"> </entity_mention> <entity_mention ID="M005" TYPE = "PRO" string = "its"> </entity_mention> <entity_mention ID="M006" TYPE = "NAME" string = "Nats"> </entity_mention> </entity>

27(130)

Annotation Tools (1): GATE

28(130)

Annotation Tools (2): Alembic

29(130)

Performance Evaluation

• Evaluation metric – mathematically defines how to measure the system’s performance against human-annotated gold standard

• Scoring program – implements the metric and provides performance measures – For each document and over the entire

corpus– For each type of NE

30(130)

Evaluation Metrics• Most common are “Precision” and “Recall”• Precision = correct answers/answers produced • Recall = correct answers/total possible correct

answers• Trade-off between precision and recall • F-Measure = (β2 + 1)PR / β2R + P

[van Rijsbergen 75]• β reflects the weighting between precision and recall,

typically β=1• Some tasks sometimes use other metrics, e.g.:

– false positives (not sensitive to doc richness)– cost-based (good for application-specific adjustment)

31(130)

The Evaluation Metric (2)• We may also want to take account of

partially correct answers:• Precision =

Correct + ½ Partially correctCorrect + Incorrect + Partial

• Recall = Correct + ½ Partially correctCorrect + Missing + Partial

• Why: NE boundaries are often misplaced, sosome partially correct results

32(130)

The GATE Evaluation Tool

33(130)

Corpus-level Regression Testing

• Need to track system’s performance over time

• When a change is made we want to know implications over whole corpus

• Why: because an improvement in one case can lead to problems in others

• GATE offers automated tool to help with the NE development task over time

34(130)

Regression Testing (2)At corpus level – GATE’s corpus benchmark tool – tracking system’s performance over time

35(130)

SW IE Evaluation tasks• Detection of entities and events, given a target

ontology of the domain.• Disambiguation of the entities and events from the

documents with respect to instances in the given ontology. For example, measuring whether the IE correctly disambiguated “Cambridge” in the text to the correct instance: Cambridge, UK vs Cambridge, MA.

• Decision when a new instance needs to be added to the ontology, because the text contains a new instance, that does not already exist in the ontology.

36(130)

Challenge:Evaluating Richer NE Tagging

• Need for new metrics when evaluating hierarchy/ontology-based NE tagging

• Need to take into account distance in the hierarchy

• Tagging a company as a charity is less wrong than tagging it as a person

37(130)









38(130)

Two kinds of IE approaches

Knowledge Engineering

• rule based • developed by experienced

language engineers • make use of human

intuition • requires only small amount

of training data• development could be very

time consuming • some changes may be

hard to accommodate

Learning Systems

• use statistics or other machine learning

• developers do not need LE expertise

• requires large amounts of annotated training data

• some changes may require re-annotation of the entire training corpus

• annotators are cheap (but you get what you pay for!)

39(130)

A) NE Baseline: list lookup approach

• System that recognises only entities stored in its lists (gazetteers).

• Advantages - Simple, fast, language independent, easy to retarget (just create lists)

• Disadvantages – impossible to enumerate all names, collection and maintenance of lists, cannot deal with name variants, cannot resolve ambiguity

40(130)

B) Shallow parsing approach using internal structure

• Internal evidence – names often have internal structure. These components can be either stored or guessed, e.g. location:

• Cap. Word + {City, Forest, Center, River}

• e.g. Sherwood Forest

• Cap. Word + {Street, Boulevard, Avenue, Crescent, Road}

• e.g. Portobello Street

41(130)

Problems ...

• Ambiguously capitalised words (first word in sentence)[All American Bank] vs. All [State Police]

• Semantic ambiguity"John F. Kennedy" = airport (location) "Philip Morris" = organisation

• Structural ambiguity [Cable and Wireless] vs. [Microsoft] and [Dell];[Center for Computational Linguistics] vs. message from [City Hospital] for [John Smith]

42(130)

C) Shallow parsing with context

• Use of context-based patterns is helpful in ambiguous cases

• "David Walton" and "Goldman Sachs" are indistinguishable

• But with the phrase "David Walton of Goldman Sachs" and the Person entity "David Walton" recognised, we can use the pattern "[Person] of [Organization]" to identify "Goldman Sachs“ correctly.

43(130)

Examples of context patterns

• [PERSON] earns [MONEY]• [PERSON] joined [ORGANIZATION]• [PERSON] left [ORGANIZATION]• [PERSON] joined [ORGANIZATION] as [JOBTITLE]• [ORGANIZATION]'s [JOBTITLE] [PERSON]• [ORGANIZATION] [JOBTITLE] [PERSON]• the [ORGANIZATION] [JOBTITLE]• part of the [ORGANIZATION]• [ORGANIZATION] headquarters in [LOCATION]• price of [ORGANIZATION]• sale of [ORGANIZATION]• investors in [ORGANIZATION]• [ORGANIZATION] is worth [MONEY]• [JOBTITLE] [PERSON]• [PERSON], [JOBTITLE]

44(130)

Example Rule-based System - ANNIE

• ANNIE – A Nearly-New IE system• A version distributed as part of GATE• GATE automatically deals with document

formats, saving of results, evaluation, and visualisation of results for debugging

• GATE has a finite-state pattern-action rule language, used by ANNIE

• A reusable and easily extendable set of components

45(130)

NE Components

46(130)

Gazetteer lists for rule-based NE• Needed to store the indicator strings for

the internal structure and context rules:• Internal location indicators – e.g., {river,

mountain, forest} for natural locations; {street, road, crescent, place, square, …}for address locations

• Internal organisation indicators – e.g., company designators {GmbH, Ltd, Inc, …}

• Produces Lookup results of the given kind

47(130)

The Named Entity Transducers• Phases run sequentially and constitute a

cascade of FSTs over the pre-processing results• Hand-coded rules applied to annotations to

identify NEs • Annotations from format analysis, tokeniser,

sentence splitter, POS tagger, and gazetteer modules

• Use contextual information • Finds person names, locations, organisations,

dates, addresses.

48(130)

NE Rule in JAPEJAPE: a Java Annotation Patterns Engine• Light, robust regular-expression-based processing • Cascaded finite state transduction • Low-overhead development of new components• Simplifies multi-phase regex processing Rule: Company1 Priority: 25 ( ( {Token.orthography == upperInitial} )+ //from tokeniser {Lookup.kind == companyDesignator} //from gazetteer lists ):match --> :match.NamedEntity = { kind=company, rule=“Company1” }

49(130)

Nam

ed E

ntiti

es in

GA

TE

50(130)

Using co-reference to classify ambiguous NEs

• Orthographic co-reference module matches proper names in a document

• Improves NE results by assigning entity type to previously unclassified names, based on relations with classified NEs

• May not reclassify already classified entities• Classification of unknown entities very useful for

surnames which match a full name, or abbreviations, e.g. [Bonfield] will match [Sir Peter Bonfield]; [International Business Machines Ltd.] will match [IBM]

51(130)

Named Entity Coreference

52(130)









53(130)

Machine Learning Approaches• Approaches:

– Train ML models on manually annotated text– Mixed initiative learning

• Used for producing training data• Used for producing working systems

• ML Methods– Symbolic learning: rules/decision trees

induction– Statistical models: HMMs, Bayesian methods,

Maximum Entropy

54(130)

ML Terminology

• Instances (tokens, entities)Occurrences of a phenomenon

• Attributes (features)Characteristics of the instances

• ClassesSets of similar instances

55(130)

Methodology

• The task can be broken into several subtasks (that can use different methods):– Boundary detection– Entity classification into NE types– Different models for different entity types

• Several models can be used in competition.– Some algorithms perform better on little data while

others are better when more training is available

56(130)

Methodology (2)Boundaries (and entity types) notations

– S(-XXX), E(-XXX)<S-ORG/>U.N.<E-ORG/> official <S-PER/>Ekeus<E-PER/> heads for <S-LOC/>Baghdad<E-LOC/>.

– IOB notation (Inside, Outside, Beginning_of)U.N. I-ORGofficial OEkeus I-PERheads Ofor OBaghdad I-LOC

. O

– Translations between the two conventions are straight-forward

57(130)

Features

• Feature selection – the most difficult part• Some automatic scoring methods can be used

• Document structure– Original markup

– Paragraph/sentence structure

• Surface features– Token length

– Capitalisation

– Token type (word, punctuation, symbol)

• Linguistic features– POS– Morphology– Syntax– Lexicon data

• Semantic features– Ontological class

• ETC– …

58(130)

Mixed Initiative Learning

• Human – computer interaction

• Speeds up the creation of training data

• Can be used for corpus/system creation

• Example implementations:– Alembic [Day et al’97] and later– Amilcare [Ciravegna’03]

59(130)

Mixed Initiative Learning (2)

P>t 1

P>t 2

User annotates System learns

System annotates User corrects System learns

System annotates

60(130)

Example 1: Alembic, Day et al 1997

• Mixed initiative approach implemented in Alembic Workbench

• Bootstrapping procedure – use already tagged data to pre-annotate new documents

• Transforms the process from tagging to review• Finally, the trained system can be used on its

own

61(130)

Mixed-Initiative text annotation

User can also edit theinduced rulesand writenew ones

Brill-style learning – generate-and-test

62(130)

Considerations• Too high recall and the human might

become over-reliant on the system annotations

• Too high precision might have similar effect

• “Theory-creep” – the choices of the human annotator are increasingly influenced by the machine’s and might deviate from the task definition – measure inter-annotator agreement

63(130)

Example 2: Amilcare & Melita

• Amilcare: rule-learning algorithm– Tagging rules – learn to insert tags in the text,

given training examples– Correction rules – learn to move already

inserted tags to their correct place in the text• Learns begin and end tags independently • Melita support adaptive IE• Applied in SemWeb context (see below)• Being extended as part of the EU-funded

DOT.KOM project towards KM andSemWeb applications

64(130)

Comparison of Alembic & Melita• The life cycle of user tagging, the machine

learning, then making suggestions which user corrects, is very similar in Melita and Alembic

• Alembic is more oriented towards NLP developers, while Melita – more towards end-users

• Melita considers timeliness and intrusiveness as criteria, while Alembic does not (possibly due to performance bottleneck from old hardware)

• Both acknowledge but do not address problems with over-reliance on machine annotations

• From ML perspective the two are very similar – rule-learning

65(130)

Eg. 3: GATE Machine Learning support• Uses classification.

[Attr1, Attr2, Attr3, … Attrn] Class• Classifies annotations.

(Documents can be classified as well using a 1-to1 relation with annotations.)

• Annotations of a particular type are selected as instances.• Attributes refer to features of the instance annotations or

their context.• Generic implementation for attribute collection – can be

linked to any ML engine.• ML engines currently integrated: WEKA and Ontotext’s

HMM.

66(130)

Implementation

Machine Learning PR in GATE.Has two functioning modes:

– training– application

Uses an XML file for configuration:<?xml version="1.0" encoding="windows-1252"?><ML-CONFIG>

<DATASET> … </DATASET><ENGINE>…</ENGINE>

<ML-CONFIG>

67(130)

Attributes Collection

Instances type: Token

68(130)

GATE MLLibrary

Dataflow

NLP Pipeline

TokeniserGazetteer

POS TaggerLexicon LookupSemantic Tagger

etc…

MachineLearningEngine

Fe

atu

reC

olle

cti

on

Re

su

lts

Co

nv

erte

r

Annotateddocuments

EngineInterface

Plain textdocuments

69(130)









70(130)

Towards Semantic Tagging of Entities

• The MUC NE task tags selected segments of text whenever that text represents the name of an entity.

• Semantic tagging - view as mentions of the underlying instances from the ontology

• Identify which mentions in the text refer to which instances in the ontology, e.g., – Tony Blair, Mr. Blair, he, the prime minister, he– Gordon Brown, he, Mr. Brown, the chancellor

71(130)

Tasks

• Identify entity mentions in the text

• Reference disambiguation– Add new instances if needed– Disambiguate wrt instances in the ontology

• Identify instances of attributes and relations– take into account what are allowed given the

ontology, using domain&range as constraints

72(130)

ExampleXYZ was established on 03 November 1978 in London. It opened a plant in Bulgaria in …

Ontology & KB

Company

type

HQ

establOn

City Country

Location

partOf

type

type type

“03/11/1978”

XYZ

London

UK Bulgaria

HQpartOf

73(130)

EntityPerson

…

Job-title

president

chancellorminister

…

G.Brown

“Gordon Brown met George Bush during his two day visit.

Classes, instances & metadata

Classes+instances before

Bush

<metadata> <DOC-ID>http://… 1.html</DOC-ID> <Annotation> <s_offset> 0 </s_offset> <e_offset> 12 </e_offset> <string>Gordon Brown</string>

<class>…#Person</class> <inst>…#Person12345</inst>

</Annotation> <Annotation> <s_offset> 18 </s_offset> <e_offset> 32 </e_offset> <string>George Bush</string>


</Annotation></metadata>

Classes+instances after

74(130)

EntityPerson

…

Job-title

president

chancellorminister

…

T. Blair

“Gordon Brown met Tony Blair to discuss the university tuition fees.

Classes, instances & metadata (2)

Classes+instances before

G. Brown

<metadata> <DOC-ID>http://… 2.html</DOC-ID> <Annotation> <s_offset> 0 </s_offset> <e_offset> 12 </e_offset> <string>Gordon Brown</string>


</Annotation> <Annotation> <s_offset> 18 </s_offset> <e_offset> 30 </e_offset> <string>Tony Blair</string>



Classes+instances after

G. Bush

75(130)

Why not put metadata in ontologies?

• Can be encoded in RDF/OWL, etc. but does it need to be put as instances in the ontology?

• Typically we do not need to reason with it– Reasoning happens in the ontology when the new instances of

classes and properties are added, but the metadata statements are different from them, they only refer to them

• A lot more metadata than instances– Millions of metadata statements, thousands of instances, hundreds

of concepts

• Different access required:– By offset (give me all metadata of the first paragraph)– Efficient metadata-wide statistics based on strings – not an

operation that people would do on other concepts– Mixing with keyword-based search using IR-style indexing

76(130)

Metadata Creation with IE

• Semantic tagging creates metadata• Stand-off or part of document• Semi-automatic

– One view (given by the user, one ontology)– More reliable

• Automatic metadata creation – Many views – change with ontology, re-train IE engine

for each ontology– Always up to date, if ontology changes– Less reliable

77(130)

Problems with “traditional” IE for metadata creation

• S-CREAM – Semi-automatic CREAtion of Metadata [Handschuh et al’02]

• Semantic tags from IE need to be mapped to instances of concepts, attributes or relations

• Most ML-based IE systems do not deal well with relations, mainly entities

• Amilcare does not handle anaphora resolution, GATE has such component but not used here

• Implemented a discourse model with logical rules – LASIE used discourse model with domain

ontology – problem is robustness and domain portability

78(130)

Example[Handschuh et al’02] S-CREAM, EKAW’02

79(130)

S-CREAM: Discourse Rules

• Rules to attach instances only when the ontology allows that (e.g., prices)

• Attach tag values to the nearest preceding compatible entity (e.g., prices and rooms)

• Create a complex object between two concept instances if they are adjacent (e.g., rate – number followed by currency)

• Experienced users can write new rules

80(130)

Challenges for IE for SemWeb

• Portability – different and changing ontologies• Different text types – structured, free, etc.• Utilise ontology information where available• Train from small amount of annotated text• Output results wrt the given ontology

– bridge the gap demonstrated in S-CREAM

• Learn/Model at the right level – ontologies are hierarchical and data will get sparser

the lower we go[DOT.KOM http://nlp.shef.ac.uk/dot.kom/]

81(130)






5. Semantic Tagging– Using “traditional” IE– Ontology-based IE (OBIE)– Platforms for large-scale processing



82(130)

Eg. 1: GATE & metadata extraction

• Combines learning and rule-based methods

• Allows combination of IE and IR

• Enables use of large-scale linguistic resources for IE, such as WordNet

• Supports ontologies as part of IE applications - Ontology-Based IE (OBIE)

83(130)

Ontology Management in GATE

84(130)

Information Retrieval

Currently based on the Lucene IR engine – useful for combining semantic and keyword-based search

85(130)

Wor

dNet

sup

port

86(130)

Populating Ontologies with IE

87(130)

Example 2: OBIE in h-TechSight

• hTechSight project – using Ontology-Based IE for semantic tagging of job adverts, news and reports in chemical engineering domain

• Aim is to track technological change over time through terminological analysis

• Fundamental to the application is a domain-specific ontology

• Terminological gazetteer lists are linked to classes in the ontology

• Rules classify the mentions in the text wrt the domain ontology

• Annotations output into a database or as an ontology

88(130)

89(130)

90(130)

Exported Database

91(130)









92(130)

Platforms for Large-Scale Metadata Creation

• Allow use of corpus-wide statistics to improve metadata quality, e.g., disambiguation

• Automated alias discovery • Generate SemWeb output (RDF, OWL)• Stand-off storage and indexing of metadata• Use large instance bases to disambiguate to• Ontology servers for reasoning and access• Architecture elements:

– Crawler, onto storage, doc indexing, query, annotators– Apps: sem browsers, authoring tools, etc.

93(130)

Example 1: SemTag • Lookup of all instances from the ontology (TAP) –

65K instances• Disambiguate the occurrences as:

– One of those in the taxonomy– Not present in the taxonomy

• Not very high ambiguity of instances with the same label in TAP – concentrate on the second problem

• Use bag-of-words approach for disambiguation• 3 people evaluated 200 labels in context – agreed

on only 68.5% - metonymy• Placing labels in the taxonomy is hard

Dill et al, SemTag and Seeker. WWW’03

94(130)

Seeker • High-performance distributed infrastructure

• 128 dual-processor machines with separate ½ terabyte of storage

• Each node runs approx. 200 documents per sec.

• Service-oriented architecture – Vinci (SOAP)

Dill et al, SemTag and Seeker. WWW’03

95(130)

Example 2: OBIE in KIM

Popov et al. KIM. ISWC’03

• The ontology (KIMO) and 86K/200K instances KB• High ambiguity of instances with the same label –

need for disambiguation step• Lookup phase marks mentions from the ontology• Combined with rule-based IE system to recognise

new instances of concepts and relations• Special KB enrichment stage where some of these

new instances are added to the KB• Disambiguation uses an Entity Ranking algorithm,

i.e., priority ordering of entities with the same label based on corpus statistics (e.g., Paris)

96(130)

OBIE in KIM (2)

Popov et al. KIM. ISWC’03

97(130)

Comparison between SemTag & KIM

• SemTag only aims for accuracy (precision) of classification of the annotated entities

• KIM also aims for coverage (recall) – whether all possible mentions of entities were found

• Trade-off – sometimes finding some is enough

• SemTag does not attempt to discover and expand the KB with new instances (e.g., new company) – the reason why KIM uses IE, not simple KB lookup

• i.e. OBIE is often needed for ontology population, not just metadata creation

98(130)

Two Annotation Scenarios (1)

• Getting the instances and the relations between them is enough, maybe not all mentions in the text are covered, but compensated by giving access to this info from the annotated text

99(130)

EntityPerson

…

Job-title

president

chancellorminister

…

G.Brown

“Gordon Brown met president Bush during his two day visit. Afterwards George Bush said…

Example

EntityPerson

…

Job-title

president

chancellorminister

…

G.BrownBush

BenchmarkThe system

Score: 100%

Bush

100(130)

Two Annotation Scenarios (2)

• Exhaustive annotation is required, so all occurrences of all instances and relations are needed

• Allows sentence and paragraph-level exploration, rather than document-level as in the previous scenario

• Harder to achieve

• Distinction between these scenarios needs to be made in the metadata annotation tools/KM tools using IE

101(130)

“Gordon Brown met president Bush during his two day visit. Afterwards George Bush said…

Example

<metadata> <Annotation> <s_offset> 0 </s_offset> <e_offset> 12 </e_offset>


</Annotation> <Annotation> <s_offset> 18 </s_offset> <e_offset> 32 </e_offset>





<metadata> <Annotation> <s_offset> 0 </s_offset> <e_offset> 12 </e_offset>





Score: 66%

102(130)

Eg. 3: SWAN: a Semantic Web Annotator

• Collaboration between DERI/NUIG, OntoText and USFD, hosted at DERI

• GATE + KIM + SECO

• Custom indexing of news or other web fractions• Quantitative media reporting• Annotated web workbench service• Custom knowledge services

• Demo and poster at ESWS

103(130)

SWAN Logical ArchitectureWeb

ServiceUsers

UIUsers

FocussedcrawlingFocussed

crawlingFocussedcrawlingFocussed

crawlingFocussedcrawling

FocussedcrawlingFocussed

crawlingFocussedcrawlingFocussed

crawlingIE(32 bit)

IE(64 bit)

Web UI,Web services

Knowledgebase

(Sesame)

Annotation(Oracle)

104(130)

Cluster Controller

105(130)

Semantic Reference Disambiguation

• Possible approaches:– Vector-space models – compare context

similarity – runs over a corpus• SemTag• Bagga’s cross-document coreference work

– Communities of practise approach from KM– Identity criteria from the ontology based on

properties, e.g., date_of_birth, name

106(130)

Why disambiguation is hard – not all knowledge is explicit in text

Paris fashion week underway as cancellations continueBy Jo Johnson and Holly Finn - Oct 07 2001 18:48:17 (FT)

Even as Paris fashion week opened at the weekend, the cancellations and reschedulings were still trickling in over the fax machines: Loewe, the leather specialists owned by LVMH empire, is not showing, Cerruti, the Italian tailor,is downscaling to private viewings, Helmut Lang, master of the sharp suit, is cancelling his catwalk.

The Oscar de la Renta show, for example, which had been planned for September 11th in New York, and which might easily enough have moved over to Paris instead, is not on the schedule. When the Dominican Republic-born designer consulted America Vogue's influential editor, Anna Wintour, she reportedly told him it would be unpatriotic to decamp.

107(130)









108(130)

Natural Language Generation

• NLG is:– “subfield of AI and CL that is concerned with

the construction of computer systems that can produce understandable texts in English or other human languages from some underlying linguistic representation of information” [Reiter&Dale’97]

– NLG techniques are applied also for producing speech, e.g., in speech dialogue systems

109(130)

Natural Language Generation

Text

Ontology/KB/Database Lexicons +Grammars

110(130)

Requirements Analysis

• Create a corpus of target texts and (if possible) their input representations

• Analyse the information content– Unchanging texts: thank you, hello, etc.– Directly available data: timetable of buses – Computable data: number of buses – Unavailable data: not in the system’s KB/DB

111(130)

NLG Tasks

1. Content determination

2. Discourse planning

3. Sentence aggregation

4. Lexicalisation

5. Referring expression generation

6. Linguistic realisation

112(130)

Content determination

• What information to include in the text – filtering and summarising input data into a formal knowledge representation

• Application dependent

• Example [ project: AKT

start_date: October-2000

end_date: October-2006

participants: {A,E,OU,So,Sh}]

113(130)

Discourse Planning

• Determine ordering and structure over the knowledge to be generated

• Theories of discourse – how texts are structured

• Influences text readability

• Result: tree structure imposing ordering over the predicates and possibly providing discourse relations

114(130)

Example

Root

Project participants

[project:AKT duration: 6 yrs]

[project: AKT participant:Shef]

Participant descr.

[univ: ShefWeb-page: URL]

Participant descr.

[project: AKT participant:OU]

…

SEQUENCE

LIST

ELABORATION ELABORATION

…

115(130)

Planning-Based Approaches

• Use AI-style planners (e.g., [Moore & Paris 93]– Discourse relations (e.g., ELABORATION) are

encoded as planning operators– Preconditions specify when the relation can

apply– Planning starts from a top-level goal, e.g.,

define-project(X)

• Computationally expensive and require a lot of knowledge – problem for real-world systems

116(130)

Schema-Based Approaches

• Capture typical text structuring patterns in templates (derived from corpus), e.g., [McKeown 85]

• Typically implemented as RTN• Variety comes from different available

knowledge for each entity• Reusable ones available: Exemplars • Example:

Describe-Project-Schema -> Sequence([duration], ProjParticipants-

Schema)

117(130)

Sentence Aggregation

• Determine which predicates should be grouped together in sentences

• Less understood process• Default: each predicate can be expressed as a

sentence, so optional step• SPOT: trainable planner• Example:

AKT is a 6-year project with 5 participants: • Sheffield (URL)• OU …

118(130)

Lexicalisation

• Choosing words and phrases to express the concepts and relations in predicates

• Trivial solution: 1-1 mapping between concepts/relations and lexical entries

• Variation is useful to avoid repetitiveness and also convey pragmatic distinctions (e.g. formality)

119(130)

Referring Expression Generation

• Choose pronouns/phrases to refer to the entities in the text

• Example: he vs Mr Smith vs John Smith, the president of XXX Corp.

• Depends on what is previously said– He is only appropriate if the person is already

introduced in the text

120(130)

Linguistic Realisation

• Use grammar to generate text which is grammatical, i.e., syntactically and morphologically correct

• Domain-independent

• Reusable components are available – e.g., RealPro, FUF/SURGE

• Example: – Morphology: participant -> participants– Syntactic agreement: AKT starts on …

121(130)

Example: a GATE-based generator

• Input– The MIAKT ontology– The RDF file for the given case– The MIAKT lexicon

• Output– GATE document with the generated

text

122(130)

Lexicalising Concepts and Instances

123(130)

Example RDF Input<rdf:Description rdf:about='c:\breast_cancer_ontology.daml#01401_patient'>

<rdf:type rdf:resource='c:\breast_cancer_ontology.daml#Patient'/><NS2:has_age>68</NS2:has_age><NS2:involved_in_ta rdf:resource='c:\breast_cancer_ontology.daml#ta-soton-1069861276136'/>

</rdf:Description><rdf:Description rdf:about='c:\breast_cancer_ontology.daml#01401_mammography'>

<rdf:type rdf:resource='c:\breast_cancer_ontology.daml#Mammography'/><NS2:carried_out_on rdf:resource='c:\breast_cancer_ontology.daml#01401_patient'/><NS2:has_date>22 9 1995</NS2:has_date><NS2:produce_result rdf:resource='c:\breast_cancer_ontology.daml#image_01401_right_cc'/>

</rdf:Description><rdf:Description rdf:about='c:\breast_cancer_ontology.daml#image_01401_right_cc'>

<NS2:image_file>cancer/case0140/C_0140_1.RIGHT_CC.LJPEG</NS2:image_file><rdf:type rdf:resource='c:\breast_cancer_ontology.daml#Right_CC_Image'/><NS2:has_lateral rdf:resource='c:\breast_cancer_ontology.daml#lateral_right'/><NS2:view_of_image rdf:resource='c:\breast_cancer_ontology.daml#craniocaudal_view'/><NS2:contains_entity rdf:resource='c:\breast_cancer_ontology.daml#01401_right_cc_abnor_1'/>

</rdf:Description><rdf:Description rdf:about='c:\breast_cancer_ontology.daml#01401_right_cc_abnor_1'>

<rdf:type rdf:resource='c:\breast_cancer_ontology.daml#Abnormality'/><NS2:is_finding rdf:resource='c:\breast_cancer_ontology.daml#mass_01401_right_cc_abnor_1'/><NS2:has_morph_feature rdf:resource='c:\breast_cancer_ontology.daml#shape_mammo_round'/><NS2:has_morph_feature rdf:resource='c:\breast_cancer_ontology.daml#margin_mammo_microlobulated'/><NS2:has_overall_impression rdf:resource='c:\breast_cancer_ontology.daml#assessment_probably_malignant'/>

</rdf:Description>

124(130)

CASE0140.RDF

The 68 years old patient is involved in a triple assessment procedure. The triple assessment procedure contains a mammography exam. The mammography exam is carried out on the patient on 22 9 1995. The mammography exam produced a right CC image. The right CC image contains an abnormality and it has a right lateral side and a craniocaudal view. The abnormality has a mass, a microlobulated margin , a round shape, and a probably malignant assessment.

125(130)

Further Reading on IE for SemWeb• Requirements for Information Extraction for Knowledge Management.

http://nlp.shef.ac.uk/dot.kom/publications.html • Information Extraction as a Semantic Web Technology: Requirements

and Promises. Adaptive Text Extraction and Mining workshop, 2003. • A. Kiryakov, B. Popov, et al. Semantic Annotation, Indexing, and

Retrieval. 2nd International Semantic Web Conference (ISWC2003), http://www.ontotext.com/publications/index.html#KiryakovEtAl2003

• S. Handschuh, S. Staab, R. Volz: http://www.aifb.uni-karlsruhe.de/WBS/sha/papers/p273_handschuh.pdf. On Deep Annotation. WWW’03.

• S. Dill, N. Eiron, et al: http://www.tomkinshome.com/papers/2Web/semtag.pdf . SemTag and Seeker: Bootstrapping the semantic web via automated semantic annotation. WWW’03.

• E. Motta, M. Vargas-Vera, et al: MnM: Ontology Driven Semi-Automatic and Automatic Support for Semantic Markup. : Knowledge Engineering and Knowledge Management (Ontologies and the Semantic Web), (EKAW02), http://www.aktors.org/publications/selected-papers/06.pdf

• K. Bontcheva, A. Kiryakov, H. Cunningham, B. Popov. M. Dimitrov. Semantic Web Enabled, Open Source Language Technology. Language Technology and the Semantic Web, Workshop on NLP and XML (NLPXML-2003). http://www.gate.ac.uk/sale/eacl03-semweb/bontcheva-etal-final.pdf

• Handschuh, Staab, Ciravegna. S-CREAM - Semi-automatic CREAtion of Metadata (2002) http://citeseer.nj.nec.com/529793.html

126(130)

Further Reading on “traditional” IE• [Day et al’97] D. Day, J. Aberdeen, L. Hirschman, R. Kozierok, P. Robinson, and M.

Vilain. Mixed-Initiative Development of Language Processing Systems. In Proceedings of the Fifth Conference on Applied Natural Language Processing (ANLP’97). 1997.

• [Ciravegna’02] F. Ciravegna, A. Dingli, D. Petrelli, Y. Wilks: User-System Cooperation in Document Annotation based on Information Extraction. Knowledge Engineering and Knowledge Management (Ontologies and the Semantic Web), (EKAW02), 2002.

• N. Kushmerick, B. Thomas. Adaptive information extraction: Core technologies for information agents (2002). http://citeseer.nj.nec.com/kushmerick02adaptive.html

• H. Cunningham, D. Maynard, K. Bontcheva, V. Tablan. GATE: A Framework and Graphical Development Environment for Robust NLP Tools and Applications. 40th Anniversary Meeting of the Association for Computational Linguistics (ACL'02). 2002.

• D.Maynard, K. Bontcheva and H. Cunningham. Towards a semantic extraction of named entities. Recent Advances in Natural Language Processing, Bulgaria, 2003.

• Califf and Mooney: Relational Learning of Pattern Matching Rules for Information Extraction http://citeseer.nj.nec.com/6804.html

• Borthwick. A. A Maximum Entropy Approach to Named Entity Recognition.PhD Dissertation. 1999

• Bikel D., Schwarta R., Weischedel. R. An algorithm that learns what’s in a name. Machine Learning 34, pp.211-231, 1999

• Riloff, E. (1996) "Automatically Generating Extraction Patterns from Untagged Text" Proceedings of the Thirteenth National Conference on Artificial Intelligence (AAAI-96) , 1996, pp. 1044-1049. http://www.cs.utah.edu/%7Eriloff/psfiles/aaai96.pdf

• Daelemans W. and Hoste V. Evaluation of Machine Learning Methods for Natural Language Processing Tasks. In LREC 2002 Third International Conference on Language Resources and Evaluation, pages 755–760

http://citeseer.nj.nec.com/kushmerick02adaptive.html

http://citeseer.nj.nec.com/6804.html

http://www.cs.utah.edu/%7Eriloff/psfiles/aaai96.pdf

127(130)

Further Reading on “traditional” IE• Black W.J., Rinaldi F., Mowatt D. Facile: Description of the NE System Used

For MUC-7. Proceedings of 7th Message Understanding Conference, Fairfax, VA, 19 April - 1 May, 1998.

• Collins M., Singer Y. Unsupervised models for named entity classificationIn Proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, 1999

• Collins M. Ranking Algorithms for Named-Entity Extraction: Boosting and the Voted Perceptron. Proceedings of the 40th Annual Meeting of the ACL, Philadelphia, pp. 489-496, July 2002 Gotoh Y., Renals S. Information extraction from broadcast news, Philosophical Transactions of the Royal Society of London, series A: Mathematical, Physical and Engineering Sciences, 2000.

• Grishman R. The NYU System for MUC-6 or Where's the Syntax? Proceedings of the MUC-6 workshop, Washington. November 1995.

• Krupka G. R., Hausman K. IsoQuest Inc.: Description of the NetOwlTM Extractor System as Used for MUC-7. Proceedings of 7th Message Understanding Conference, Fairfax, VA, 19 April - 1 May, 1998.

• McDonald D. Internal and External Evidence in the Identification and Semantic Categorization of Proper Names. In B.Boguraev and J. Pustejovsky editors: Corpus Processing for Lexical Acquisition. Pages21-39. MIT Press. Cambridge, MA. 1996

• Mikheev A., Grover C. and Moens M. Description of the LTG System Used for MUC-7. Proceedings of 7th Message Understanding Conference, Fairfax, VA, 19 April - 1 May, 1998

• Miller S., Crystal M., et al. BBN: Description of the SIFT System as Used for MUC-7. Proceedings of 7th Message Understanding Conference, Fairfax, VA, 19 April - 1 May, 1998

128(130)

Further Reading on multilingual IE• Palmer D., Day D.S. A Statistical Profile of the Named Entity Task.

Proceedings of the Fifth Conference on Applied Natural Language Processing, Washington, D.C., March 31- April 3, 1997.

• Sekine S., Grishman R. and Shinou H. A decision tree method for finding and classifying names in Japanese texts. Proceedings of the Sixth Workshop on Very Large Corpora, Montreal, Canada, 1998

• Sun J., Gao J.F., Zhang L., Zhou M., Huang C.N. Chinese Named Entity Identification Using Class-based Language Model. In proceeding of the 19th International Conference on Computational Linguistics (COLING2002), pp.967-973, 2002.

• Takeuchi K., Collier N. Use of Support Vector Machines in Extended Named Entity Recognition. The 6th Conference on Natural Language Learning. 2002

• D.Maynard, K. Bontcheva and H. Cunningham. Towards a semantic extraction of named entities. Recent Advances in Natural Language Processing, Bulgaria, 2003.

• M. M. Wood and S. J. Lydon and V. Tablan and D. Maynard and H. Cunningham. Using parallel texts to improve recall in IE. Recent Advances in Natural Language Processing, Bulgaria, 2003.

• D.Maynard, V. Tablan and H. Cunningham. NE recognition without training data on a language you don't speak. ACL Workshop on Multilingual and Mixed-language Named Entity Recognition: Combining Statistical and Symbolic Models, Sapporo, Japan, 2003.

129(130)

Further Reading on multilingual IE• H. Saggion, H. Cunningham, K. Bontcheva, D. Maynard, O. Hamza, Y.

Wilks. Multimedia Indexing through Multisource and Multilingual Information Extraction; the MUMIS project. Data and Knowledge Engineering, 2003.

• D. Manov and A. Kiryakov and B. Popov and K. Bontcheva and D. Maynard, H. Cunningham. Experiments with geographic knowledge for information extraction. Workshop on Analysis of Geographic References, HLT/NAACL'03, Canada, 2003.

• H. Cunningham, D. Maynard, K. Bontcheva, V. Tablan. GATE: A Framework and Graphical Development Environment for Robust NLP Tools and Applications. Proceedings of the 40th Anniversary Meeting of the Association for Computational Linguistics (ACL'02). Philadelphia, July 2002.

• H. Cunningham. GATE, a General Architecture for Text Engineering. Computers and the Humanities, volume 36, pp. 223-254, 2002.

• D. Maynard, H. Cunningham, K. Bontcheva, M. Dimitrov. Adapting A Robust Multi-Genre NE System for Automatic Content Extraction. Proc. of the 10th International Conference on Artificial Intelligence: Methodology, Systems, Applications (AIMSA 2002), 2002.

• K. Pastra, D. Maynard, H. Cunningham, O. Hamza, Y. Wilks. How feasible is the reuse of grammars for Named Entity Recognition? Language Resources and Evaluation Conference (LREC'2002), 2002.

130(130)

THANK YOU!(for not snoring)

The slides: http://gate.ac.uk/sale/talks/esws2004-tutorial.ppt

[This work has been supported by AKT (http://aktors.org/) and

SEKT (http://sekt.semanticweb.org/)]

human language technology for the semantic web //gate.ac.uk/ //nlp.shef.ac.uk/ hamish cunningham...

Documents

language loop slide

textual language

semantic web http

uksale slide

language generation

human language gartnergartner

human language technology

structured information