ESSLLI 2006 Summer school

1

• ESSLLI 2006 Summer school
• Malaga, Spain
• 31 July 11 August

2
General Comments

• PLUS
• Courses on time
• Proceedings of all courses
• Workshops
• Student sessions
• Internet connection

• MINUS
• Not well organized
• Site not updated on time
• Lunch tickets

3
Courses

• Counting Words An Introduction to Lexical
  Statistics
• Formal Ontology for Communicating Agents
  (Workshop)
• Word Sense Disambiguation
• Introduction to Corpus Resources, Annotation
  Access
• An Empirical View on Semantic Roles Within and
  Across Languages
• Approximate Reasoning for the Semantic Web

4
Counting WordsMarco Baroni and Stefan Evert

• Contents
• Introduction
• Distributions
• Zipfs Law
• The ZipfR package
• Practical Consequences and Conclusion

5
Introduction

• The frequency of words plays an important role in
  corpus linguistics.
• The study of word frequency distributions is
  called Lexical Statistics.
• It seems that word frequency distributions are
  more of interest to theoretical physicists than
  to theoretical linguists.
• This course introduces some of the empirical
  phenomena pertaining to word frequency
  distributions and the classic models that have
  been proposed to capture them.

6
DistributionsBasic Terminology

• Types distinct words
• Tokens instances of all distinct words
• Corpus size (N) number of tokens in the corpus
• Vocabulary size (V) number of types
• Frequency list list that reports the number of
  tokens of each type in the corpus
• Rank/Frequency profile replace the types with
  the frequency ranks
• Frequency Spectrum a list reporting how many
  types in a frequency list have a certain frequency

7
DistributionsExample

• Sample a b b c a a b a d
• N9, V4
• Freq. list rank/freq. prof. Freq.
  spect.

type f
a 4
b 3
c 1
d 1
rank f
1 4
2 3
3 1
4 1
f V(f)
1 2
3 1
4 1
8
DistributionsTypical frequency patterns

• Top ranks are occupied by function words (the,
  of, and..)
• Frequency decreases quite rapidly
• The lowest frequency elements are content words

9
Zipfs Law

• The frequency is a non-linear decreasing function
  of rank.
• Zipfs model f(w)C/r(w)a
• The model predicts a very rapid decrease in
  frequency among the most frequent words, which
  becomes slower as the rank grows.
• Mathematical property
• logf(w)logC-alogr(w) (Linear function)

10
Zipfs LawApplications and explanations

• Zipfian distributions are encountered in various
  phenomena
• City populations
• Incomes in economics
• Frequency of citations of scientific papers
• Visits to web sites
• Least effort principle

11
ZipfR Package

• Statistical package for modeling lexical
  distributions.
• url http//www.purl.org/stefan.evert/zipfR
• Dependencies the R package
• url http//www.r-project.org
• Binaries available for Win and MacOS.
• Source available for Linux.
• Open source, GNU Licensed project.

12
Practical Consequences and Conclusion

• The Zipfian nature of word frequency distribution
  causes data sparseness problems.
• Although V is growing with corpus size, we cannot
  use it as a measure of lexical richness when
  comparing corpora.
• Interested readers should proceed to Baayen(2001)
  for a thorough introduction to word frequency
  distributions with an emphasis to statistical
  modeling.

13
References

• Abney, Steven (1996), Statistical methods and
  linguistics. In Klavans, J. Resnik, P. (eds)
  The balancing act Combining symbolic and
  statistical approaches to language. Cambridge MA
  MIT Press, 1-23.
• Baayen, Harald (2001), Word frequency
  distributions. Dordrecht Kluwer
• Baldi, Pierre/Frasconi, Paolo/Smyth, Padhraic
  (2003), Modeling the internet and the web.
  Chichester Wiley
• Biber, Douglas/Conrad, Susan/Reppen, Randi
  (1998), Corpus linguistics. Cambridge Cambridge
  University Press
• Creutz, Mathias (2003), Unsupervised segmentation
  o words using prior distributions of morph length
  and frequency. In Proceedings of ACL 03, 280-287
• Delgaard, Peter (2002), Introductory statistics
  with R. New York Springer
• Evert, Stefan (2004), The statistics of word
  co-occurrences Word pairs and collocations.PhD
  thesis, University of Stuttgard/IMS

14
References

• Evert, Stefan/Baroni, Marco (2006), Testing the
  extrapolation quality of word frequency models.
  In Proceedings of Corpus Linguistics 2005,
  available from http//www.corpus.bham.ac.uk./PCLC
• Li, Wentian (2002), Zipfs Law everywhere. In
  Glottometrics 5, 14-21
• Manning, Christopher/Schutze, Hinrich (1999),
  Foundations of statistical natural language
  processing. Cambridge MA MIT Press
• McEnery, Tony and Andrew Wilson (2001), Corpus
  Linguistics, 2nd edition. Edinburgh Edinburgh
  University Press
• Oakes, Michael (1998), Statistics for corpus
  linguistics. Edinburgh Edinburgh University
  Press
• Sampson Geoffrey (2002), Review of Harald Baayen
  Word frequency distributions. In Computational
  Linguistics 28, 565-569
• Zipf, George Kingsley (1949), Human behavior and
  the principle of least effort. Cambridge MA
  Addison-Wesley
• Zipf, George Kingsley (1965), The psycho-biology
  of language. Cambridge MA MIT Press

15
Formal Ontology for Communicating Agents
(FOCA)Workshop

• Contents
• Introduction
• Communicative acts
• The missing ontological link
• Semantic Coordination
• A Communication Acts Ontology for Software Agents
  Interoperability
• OWL DL as a FIPA ACL content Language

16
Introduction

• Purpose of the workshop
• To gather contributions that
• Take seriously into account the ontological
  aspects of communication and interaction
• Use formal ontologies for achieving a better
  semantic coordination between interacting and
  communicating agents

17
IntroductionCommunicative acts

• According to Austin, 3 kinds of acts can be
  performed simultaneously through a single
  utterance
• Locutionary act producing noises that conform to
  a system
• Illocutionary act what is performed in saying
  something
• Perlocutionary act what is performed by saying
  something
• An important issue is the distinction between the
  last two acts.

18
IntroductionThe missing ontological link

• Ontological ingredients
• Events, states, actions, speech acts, relations,
  plans, propositions, arguments, facts,
  commitments,..
• Top-level ontologies focus on the sub-domain of
  concrete entities, like time, space,..
• There is a need for the integration of the large
  amount of the philosophical work on other domains
  like that of abstract entities.

19
IntroductionSemantic Coordination

• An important aspect of interaction and
  communication involves the management of
  ontologies.
• Scenaria identified w.r.t. semantic coordination
• With a shared pre-existing ontology
• With different ontologies but linked to a
  pre-existing common upper level ontology
• With different ontologies but mapped directly
  onto each other
• When agents are involved
• Keep static ontologies but manage a shared
  dynamic one
• Create new static ontologies through a
  negotiation phase
• Modify their ontology during the interaction
  while maintaining some kind of negotiation
  meaning

20
A Communication Acts Ontology for Software Agents
Interoperability

• Different classes of communication acts to each
  ACL.
• The use of an agreed ontology can open a
  possibility of real agents interoperation based
  on a wide agreement on some classes of
  communication acts that will serve as a bridge
  among different ACL islands
• Main design criterion follow the speech act
  theory and also embed an approach for expressing
  the semantics of the communication acts
• Use the OWL DL language

21
A Communication Acts Ontology for Software Agents
Interoperability

• Upper layer
• CommunicationAct ? ?hasSender.Actor ?
  1.hasSender ? ?hasReceiver.Actor ?
  ?hasContent.Content
• Assertive ? CommunicationAct ? ?hasContent.Proposi
  tion ? ?hasCommit.AssertiveCommitment
• Directive ? CommunicationAct ? ?hasContent.Action
  ?hasCommit.DirectiveCommitment
• Commisive ? CommunicationAct ? ?hasContent.Action
  ?hasCondition.Proposition ? ?hasCommit.CommissiveC
  ommitment
• Expressive ? CommunicationAct ?
  ?hasContent.Proposition ? ?hasState.PsyState
  ?hasCommit.ExpressiveCommitment
• Declarative ? CommunicationAct ?
  ?hasContent.Proposition

22
A Communication Acts Ontology for Software Agents
Interoperability

• The Standards Layer extends the Upper Layer with
  terms representing classes of communication acts
  of general purpose ACLs, like FIPA-ACL.
• The Applications Layer is the most specific.
  Defines communication acts classes for a specific
  application.
• Concluding Classes in the upper layer are
  considered the framework agreement for general
  communication. Classes in the standard layer
  reflect classes of communication acts that
  different standard ACLs define. Classes in the
  application layer concern the particular
  communication acts used by each agent system
  committing to the ontology.

23
References

• J. L. Austin. How to Do Things With Words. Oxford
  University Press. Oxford, 1962
• J. R. Searle. Speech Acts An Essay in the
  Philosophy of Language. Cambridge University
  Press. New York, 1969
• M. P. Singh. Agent Communication Languages
  Rethinking the Principles. IEEE Computer, vol.31,
  num.12, pp.40-47, 1998
• M. Wooldridge. Semantic Issues in the
  Verification of Agent Communication Languages.
  Journal of Autonomous Agents and Multi-Agent
  Systems, vol.3, num.1, pp.9-31, 2000
• Y. Labrou, T. Finin, Y. Pen. Agent Communication
  Languages the Current Landscape. IEEE
  Intelligent Systems, vol.14, num.2, pp.45-52,
  1999
• M. P. Singh. A Social Semantics for Agent
  Communication Languages. Issues in Agent
  Communication, pp.31-45. Spinger-Verlag, 2000
• FIPA Communicative Act Library Specification.
  Foundation For Intelligent Physical Agents, 2005.
  http//www.fipa.org/specs/fipa00037/SC00037J.html

24
References

• N. Asher and A. Lascarides. Logics of
  Conversation. Cambridge University Press, 2003
• S. Levinson. Pragmatics. Cambridge University
  Press, 1983
• J.R. Searle and D. Vanderveken. Foundations of
  illocutionary logic. Cambridge University Press,
  1975
• J.R. Searle. The Construction of Social Reality.
  Free Press, New York, 1995
• R. Stalnaker. Assertion. Syntax and Semantics,
  9315-332, 1978
• J. Ginzburg. Dynamics and the Semantics of
  Dialogue. CSLI Stanford, 1996
• H. H. Clark. Using Language. Cambridge University
  Press, 1996
• S. Carberry. Plan Recognition in Natural Language
  Dialogue. MIT Press, 1990

25
OWL DL as a FIPA ACL Content Language

• FIPA-SL content language is in general
  undecidable.
• Use OWL DL in order to enable semantic validation
  in the content of the ACL message and to separate
  speech act semantics from content semantics.
• Their ontology defines some of the FIPA
  specifications (message structure, ontology
  service, content language, communicative act lib)

26
OWL DL as a FIPA ACL Content Language

• Advantages
• Application ontologies are domain independent.
  They can be applied to a MAS in different
  domains.
• Various application ontologies in OWL DL are
  available. This shows a great potential for
  reusing already formulated ontologies.
• W3C suggests the use of OWL within agents.

27
References

• Eric Miller et al. Web Ontology Language (OWL),
  2004
• RACER Systems GmbH. The features of racerpro
  version 1.9, 2005
• Foundation for Intelligent Physical Agents. FIPA
  ACL Message Structure Specification, 2002
• Foundation for Intelligent Physical Agents. FIPA
  Ontology Service Specification, 2001
• Foundation for Intelligent Physical Agents. FIPA
  SL Content Language Specification, 2002
• Foundation for Intelligent Physical Agents. FIPA
  Communicative Act Library Specification, 2002
• Web Ontology Working Group. OWL Web Ontology
  Language Use Cases and Requirements, 2004
• Giovani Caire. JADE Introduction AAMAS 2005, 2005

28
Introduction to Corpus Resources, Annotation
AccessSabine Schulte im Walde and Heike
Zinsmeister

• Contents
• Basic definitions
• Corpora
• Annotation
• Tokenization Morpho-Syntactic Annotation

29
Introduction to Corpus Resources, Annotation
Access

• Basic Definitions
• Linguistics Characterization and explanation of
  linguistic observations
• Corpus Any collection of more than one text
• Annotation The practice of adding
  interpretative, linguistic information to an
  electronic corpus of spoken and/or written
  language

30
Corpora

• Corpora give only a partial description of a
  language
• They are incomplete
• (e.g. Brown corpus does not include vocabulary
  related to WWW and e-mail)
• They are biased
• They include ungrammatical sentences
• (e.g. typos, copy-and-paste errors, conversion
  errors)
• We have to sample a corpus according to some
  design criteria such that it is balanced and
  representative for a specific purpose

31
Annotation

• Levels
• POS tags
• Lemmata
• Senses
• Semantic roles
• Named Entities
• Topic
• Co reference

• Principles
• The raw corpus should be recoverable
• Annotation should be extricable from the corpus
• Easy access to documentation
• Annotation scheme
• How, where, by whom the annotation was applied

32
Tokenization and Morpho-Syntactic Annotation

• Tokenization divides the raw input character
  sequence of a text into sentences and the
  sentences into tokens
• Problems
• Language dependent task
• Sentence boundaries
• Numbers
• Abbreviations
• Capitalization
• Hyphenation
• Multiword expressions
• Clitics
• ? So.. We need to apply disambiguation methods

33
Tokenization and Morpho-Syntactic Annotation

• Part-Of-Speech Tagging (POS tagging) The task of
  labeling each word in a sequence of words with
  its appropriate part-of-speech.
• Performs a limited syntactic disambiguation
• Context helps to disambiguate tags
• Tagset A set of part-of-speech tags
• Classical 8 classes noun, verbs, article,
  participle, pronoun, preposition, adverb,
  conjunction

34
Tokenization and Morpho-Syntactic Annotation

• Morphology morphology is concerned with the
  inner structure of words and the formation of
  words from smaller units.
• Root the morphem of the word
• Stemming A process that strips off affixes and
  leaves the stem.
• Lemmatization A process that gives the lemma of
  a word. Includes disambiguation at the level of
  lexemes, depending on the part-of-speech.
• Co reference is the reference in one expression
  to the same referent in another expression
• Anaphora is co reference of one expression with
  its antecedent

35
References

• Tony McEnery (2003). Corpus Linguistics. In The
  Oxford Handbook of Computational Linguistics,
  pp.448-463. Oxford University Press
• Tony McEnery and Andrew Wilson (2001). Corpus
  Linguistics. 2nd edition. Edinburgh University
  Press, chapter 1
• Sue Atkins, Jeremy Clear and Nicholas Ostler
  (1992). Corpus Design Criteria. In Literary and
  Linguistic Computing, 7(1)1-16
• Nancy Ide (2004). Preparation and Analysis of
  Linguistic Corpora. In Schreibman, S., Siemens,
  R., Unsworth, J., eds. A Companion to Digital
  Humanities. Blackwell
• Geoffrey Leech (1997). Introducing Corpus
  Annotation. In Richard Garside, Geoffrey Leech
  and Tony McEnery, eds. Corpus Annotation.
  Longmanm pp.1-18
• Geoffrey Leech (2005). Adding Linguistic
  Annotation. In Developing Linguistic Corpora A
  Guide to good Practice, ed. M. Wynne. Oxford
  Oxbow Books, pp. 17-29. Available online from
  http//ahds.ac.uk./linguistic-corpora/
• Gregory Grefenstette and Pasi Tapanainen (1994)
  What is a word, what is a sentence? Problems of
  tokenization. In Proceedings of the 3rd
  Conference on Computational Lexicography and Text
  Research.

36
References

• Andrei Mikheev (2003) "Text segmentation". In
  Ruslan Mitkov, editor, "The Oxford Handbook of
  Computational Linguistics", pp. 376-394. Oxford
  University Press.
• Helmut Schmid (2007?) "Tokenizing". In Anke
  Lüdeling and Merja Kytö, editors, "Corpus
  Linguistics.
• An International Handbook. Mouton de Gruyter,
  Berlin.
• Christopher D. Manning and Hinrich Schütze
  (1999) Foundations of Statistical Natural
  Language Processing, chapter 10. MIT Press.
• Atro Voutilainen (2003) Part-of-speech
  tagging". In Ruslan Mitkov, editor, "The Oxford
  Handbook of Computational Linguistics", pp.
  219-232. Oxford University Press.
• John Carroll, Guido Minnen, and Ted Briscoe
  (1999) Corpus annotation for parser
  evaluation. In Proceedings of LINC. Bergen.
• Ruslan Mitkov, Richard Evans, Constantin Orasan,
  Catalina Barbu, Lisa Jones, and Violeta Sotirova
  (2000) Coreference and anaphora developing
  annotating tools, annotated resources and
  annotation strategies. In Proceedings of the
  Discourse, Anaphora and Reference Resolution
  Conference, pp. 49-58.
• Eva Hajicová, Jarmila Panevová, and Petr Sgall
  (2000) "Coreference in annotating a large
  corpus". In Proceedings of the 2nd International
  Conference on Language Resources and Evaluation,
  pp. 497-500.

37
Approximate Reasoning for the Semantic WebFrank
van Harmelen, Pascal Hitzler and Holger Wache

• Contents
• Semantic Web the Vision
• Ontologies
• XML
• W3C Stack
• Beyond RDF OWL
• Why Approximate Reasoning
• Reduction of use-cases to reasoning methods

38
Semantic Web the Vision

• Semantic Web Web of Data
• Set of open, stable W3C standards
• Intelligent things we cant do today
• Search engines concepts, not keywords
• Personalization
• Web Services need semantic characterizations to
  find them, to combine them
• Requirement Machine Accessible Meaning

39
Ontologies

• Ontologies ARE shared models of the world
  constructed to facilitate communication
• Ontologies ARE NOT definitive descriptions of
  what exists in the world (this is philosophy)
• Whats inside an ontology?
• Classes
• Instances
• Values
• Inheritance
• Restrictions
• Relations
• Properties
• We need a machine representation

40
XML

• What was XML again?
• ltcountry nameGreecegt
• ltcapital nameAthensgt
• ltareacodegt210lt/areacodegt
• lt/capitalgt
• lt/countrygt

• Why not use XML ??
• No agreement on
• Structure
• Is country a
• Object?
• Class?
• Attribute?
• Relation?
• What does nesting mean?
• Vocabulary
• Is country the same as nation ?

country
name
capital
Greece
name
areacode
Athens
210
41
W3C Stack

• XML
• Surface syntax, no semantics
• XML Schema
• Describes structure of XML documents
• RDF
• Datamodel for relations between things
• RDF Schema
• RDF Vocabulary Definition Language
• OWL
• A more expressive Vocabulary Definition Language

42
Beyond RDF OWL

• OWL extends RDF Schema to a full-fledged ontology
  representation language.
• Domain / range
• Cardinality
• Quantifiers
• Enumeration
• Equality
• Boolean Algebra
• Union, complement
• OWL is simply a Description Logic SHOIN(D) with
  an RDF/XML syntax.
• 3 Flavors OWL Lite, OWL DL, OWL Full

43
Why Approximate Reasoning

• Current inference is exact
• yes or now
• This was OK, because until now ontologies were
  clean
• Hand-crafted, well-designed, carefully populated,
  well maintained,
• BUT, ontologies will be sloppy
• Made by machines
• (e.g. almost subClassOf)
• Mapping ontologies is almost always messy
• (e.g. almost equal)

44
Reduction of use-cases to reasoning methods

• Realization (member of)
• Subsumption (subclass-relation)
• Mapping (similar to)
• Retrieval (has member)
• Classification (locate in hierarchy)
• GOAL
• Find approximation methods for the reasoning
  methods
• Many reasoning methods can be reduced to
  satisfiability
• GOAL find approximation methods for
  satisfiability

45
References

• Cadoli and Schaerf, 1995 Marco Cadoli and Marco
  Schaerf. Approximate inference in default
  reasoning and circumscription. Fundamenta
  Informaticae, 23123143, 1995.
• Cadoli et al., 1994 Marco Cadoli, Francesco M.
  Donini, and Marco Schaerf. Is intractability of
  non-monotonic reasoning a real drawback? In
  National Conference on Artificial Intelligence,
  pages 946951, 1994.
• Dalal, 1996a M. Dalal. Semantics of an anytime
  family of reasoners. In W. Wahlster, editor,
  Proceedings of ECAI-96, pages 360364, Budapest,
  Hungary, August 1996. John Wiley Sons LTD.
• Motik, 2006 B. Motik. Reasoning in Description
  Logics using Resolution and Deductive Databases.
  PhD thesis, Universität Karlsruhe (2006)
• Schaerf and Cadoli, 1995 Marco Schaerf and
  Marco Cadoli. Tractable reasoning via
  approximation. Artificial Intelligence,
  74249310, 1995.
• Zilberstein, 1993 S. Zilberstein. Operational
  rationality through compilation of anytime
  algorithms. PhD thesis, Computer science
  division, university of California at Berkley,
  1993.
• Zilberstein, 1996 S. Zilberstein. Using anytime
  algorithms in intelligent systems. Artificial
  Intelligence Magazine, fall7383, 1996.

46
Word Sense DisambiguationRada Mihalcea

• Outline
• Some Definitions
• Basic Approaches Intro
• Basic Approaches In more Detail
• Some Examples

47
Word Sense Disambiguation

• Word Sense Disambiguation is the problem of
  selecting a sense for a word from a set of
  predefined possibilities (Sense Inventory).
• Sense Inventory usually comes from a dictionary
• Word Sense Discrimination is the problem of
  dividing the usages of a word into different
  meanings, without regard to existing predefined
  possibilities.

48
Word Sense Disambiguation

• Knowledge-Based Disambiguation
• - Machine Readable Dictionaries (e.g. WordNet)
• - Raw Corpora (not manually annotated)
• Supervised Disambiguation
• - Manually Annotated Corpora
• - Input of the learning system is
• 1. a training set of the feature-encoded inputs
• 2. their appropriate sense label
• Unsupervised Disambiguation
• - Unlabelled corpora
• - Input of the learning system is
• 1. a training set of feature-encoded inputs
• 2. NOT their appropriate sense label

49
Word Sense Disambiguation

• Knowledge-Based Disambiguation
• Examples
• - Algorithms based on Machine Readable
• Dictionaries (e.g. Lesk alg)
• - Semantic Similarity Metrics
• - relies on semantic networks, like ontologies
• e.g. Sim(a,b) -log(Path(a,b))/2D)
• - may utilize on information content metric
• e.g. Sim(a,b) IC(LCS(a,b)), IC(a)-log(P(a))
• - Heuristic-based Methods
• e.g. identify the most often used meaning and
  use it
• by default.

50
Word Sense Disambiguation

• Knowledge-Based Disambiguation
• Examples
• disambiguate plant in plant with flower
• 1. plant, works, industrial plant
• 2. plant, flora, plant life
• Sim(plant1, flower)1.0
• Sim(plant2, flower)1.5 winner sense 2

51
Word Sense Disambiguation

• Supervised Disambiguation
• -Class of methods that induce a classifier from
  manually sense-tagged text using machine learning
  techniques (SVM, Na?ve Bayes, Neural Networks..)
• - Resources
• 1. Sense tagged text
• 2. Dictionary (source of sense inventory)
• 3. Syntactic Analysis (POS tagger, Chunker)
• Example of features of a training algorithm for
  the target word bank bank/SHORE and
  bank/FINANCE

52
Word Sense Disambiguation

• Unsupervised Disambiguation
• - Identifies patterns and divides data into
  clusters,
• where its member of a cluster has more in common
• with the members of its own class, than any other
• - Words with similar meanings tend to occur in
  similar
• contexts. So clustering is based on the context
• - Only raw text is available, no external
  resources nor
• annotations
• - Usual Approaches Agglomerative algorithm, LSA

53
Word Sense Disambiguation

• Unsupervised Disambiguation
• Examples
• - Agglomerative Clustering(McQuitty's Similarity
  Analysis)
• First Order Representation of the target word
  bank, in four sentences

Similarity Matrix and resulting clustering
54
An Empirical View on Semantic Roles Within and
Across LanguagesKatrin Erk and Sebastian Pado

• Outline
• - The problem
• - Predicate-argument structure
• - A solution
• Proposition Bank (PropBank)
• (http//www.cs.rochester.edu/gildea/PropBank/Sort
  )

55
An Empirical View on Semantic Roles Within and
Across Languages

• The problem
• - Despite of the breakthroughs in NLP based on
  statistical
• methods and linguistic representations, accurate
  information
• extraction was out of reach
• - A critical element was missing accurate
  predicateargument
• structure
• - The most important factor for improved quality
  in language
• translation is accurate predicate-argument
  structure
• - Complete grammatical parse and vocabulary
  coverage are
• not enough.
• - Knowledge of the proper constituents of verb
  arguments is
• not enough. Their proper position is very
  important

56
An Empirical View on Semantic Roles Within and
Across Languages

• Predicate-argument structure
• - Example
• Sentence John broke the window
• Associated predicate-argument break(John,
  window)
• - The recognition of the structure is not a
  trivial problem
• - In natural language there are several lexical
  items referring
• to the same type of event and several syntactic
  realizations
• of the same predicate-argument relations
• - Example
• A will meet/visit/consult/debate (with) B
• A and B met/visited/consulted/debated
• There was a meeting/visit/consultation/debate
• between A and B
• A had a meeting/visit/consultation/debate with
  B

57
An Empirical View on Semantic Roles Within and
Across Languages

• A solution
• - Create a body of publicly available training
  data that explicitly annotates predicate-argument
  positions with labels.
• - Highest priority was given to
  predicate-argument structure for verbs
• - The result was the Proposition Bank (PropBank)

58
An Empirical View on Semantic Roles Within and
Across Languages

• Proposition Bank (PropBank)
• - 4000 predicates (verbs only)
• - Process
• 1. For a given predicate a survey is made of the
  its usages
• 2. The usages are divided into senses if they
  take different
• number of arguments (syntactic grounds, not
  semantic)
• 3. The expected arguments of each sense are
  numbered
• sequentially from Arg0 to Arg5
• - Example
• draw sense pull
• ... the campaign is drawing fire from the
  anti-smoking
• advocates...
• Arg0 the campaign
• Re1 drawing
• Arg1 fire
• Arg2-from anti-smoking advocates

59
An Empirical View on Semantic Roles Within and
Across Languages

• Proposition Bank (PropBank)
• - Frame Files (developed by a linguist)
• 1. Contain sense distinctions of predicates
  (previous
• slide)
• 2. Contain role sets. A role set of a verb
  lists the roles
• which seem to occur more frequently.
• - Example of role set for verb buy
• BUY
• Arg0 buyer
• Arg1 thing bought
• Arg2 seller, bought-from
• Arg3 price paid
• Arg4 putrefactive, bought-for

60
ESSLLI 2006 Summer School

• 18th European Summer School in Logic, Language
  and Information
• Thanks!