Structure of this course Semantic Theory: Lexical Semantics II • Basic Semantic Relations • WordNet Summer 2008 • Predicate-Argument Structure • PropBank and FrameNet M.Pinkal/ S. Thater • Event Semantics Semantic Theory, SS 2008 © M. Pinkal, S. Thater 2 Basic Semantic Relations Relations expressing contrast • Synonymy: – Identity of meaning (context-specific, for a given sense) • Antonymy or Contrast: • Hyponymy (inverse relation: Hypernymy): – good – bad, expensive – cheap, tall – short – the sub-/superconcept relation: – car - truck, dog - animal, kill - murder • Complementarity: • Meronymy (inverse relation: Holonymy): – man – woman, married – single, alive – dead – General part-of relation, with three (well-motivated) sub-relations: • Converseness/ inverse relation: – Physical Part – Whole relation: branch – tree – buy – sell, parent – child, taller than – shorter than – Member – Group relation: tree – forest – Matter – Object relation: wood – tree Semantic Theory, SS 2008 © M. Pinkal, S. Thater 3 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 4 WordNet Senses of car • WordNet represents a layer of the semantic lexicon of • S: (n) car, auto, automobile, machine, motorcar English as a network of semantic relations, with the hyponymy relation as its backbone. • S: (n) car, railcar, railway car, railroad car • The nodes of the semantic network are „synsets“: Sets • S: (n) car, gondola of synonymous words, which represent concepts/ word senses. • S: (n) car, elevator car • Synsets directly provide synonymy information, and • S: (n) cable car, car information about the word-concept mapping: A (orthographic) word has all those senses/ synsets as readings, of which it is a member. • In cases where no or too few synonyms are available for sense distinction, WordNet glosses and examples help to disambiguate. Semantic Theory, SS 2008 © M. Pinkal, S. Thater 5 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 6 Synsets + glosses + examples Hyponyms of motor vehicle • S: (n) motor vehicle, automotive vehicle (a self-propelled wheeled vehicle that does not run on rails) • car • direct hyponym / full hyponym – S: (n) amphibian, amphibious vehicle (a flat-bottomed motor vehicle that can travel – { car, auto, automobile, machine, motorcar } on land or water) – a motor vehicle with four wheels; usually propelled by – S: (n) bloodmobile (a motor vehicle equipped to collect blood donations) – S: (n) car, auto, automobile, machine, motorcar (a motor vehicle with four wheels; an internal combustion engine usually propelled by an internal combustion engine) "he needs a car to get to – "he needs a car to get to work" work" – S: (n) doodlebug (a small motor vehicle) – S: (n) four-wheel drive, 4WD (a motor vehicle with a four-wheel drive transmission system) – S: (n) go-kart (a small low motor vehicle with four wheels and an open framework; used for racing) – S: (n) golfcart, golf cart (a small motor vehicle in which golfers can ride between shots) – S: (n) hearse (a vehicle for carrying a coffin to a church or a cemetery; formerly drawn by horses but now usually a motor vehicle) – S: (n) motorcycle, bike (a motor vehicle with two wheels and a strong frame) – S: (n) snowplow, snowplough (a vehicle used to push snow from roads) – S: (n) truck, motortruck (an automotive vehicle suitable for hauling) Semantic Theory, SS 2008 © M. Pinkal, S. Thater 7 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 8 A small fragment of the WN graph WordNet • English WordNet is by far the largest lexical-semantic resource: – 150.000 lexical items – 120.000 synsets – 200.000 word-sense pairs • WordNet is extensively used in many Language technology applications. • Versions of WordNet currently available for about 45 languages (with large differences in coverage, design, and availability) • "GermaNet": a German WordNet version with about 100.000 lexical items. Semantic Theory, SS 2008 © M. Pinkal, S. Thater 9 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 10 Basic Uses of WordNet Information in Limitations of WordNet Language technology • Query expansion with WordNet synonyms/hyponyms • Much of WordNet information is only informally represented (in glosses and examples). • Measuring semantic distance by (normalised) path • WordNet consists of different unrelated data-bases for common length nouns, verbs, adjectives (and adverbs). - No information about • WordNet as an ontology, a database of axioms feeding cross-categorial sense distinctions. logical inference • Wide variation of granularity in different parts of WordNet. !x(family(x)"group(x)) • In general, WordNet tends to be too fine-granular (branching factor and depth of hierarchy). !x(person(x) " #y(substance_m(y,x) $ body(y)) • WordNet focusses on paratactic semantic relations. No information !x(body(x) " #y(part_m(y,x) $ leg(y)) how to build predicate-argument structure. !x(body(x) " #y(part_m(y,x) $ arm(y)) • No information about selectional constraints/ preferences. (Next week: Description Logic) • WordNet (of course) does not solve the (notoriously hard) problem of word-sense disambiguation. Semantic Theory, SS 2008 © M. Pinkal, S. Thater 11 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 12 Limitations of WordNet Extended WordNet • Much of WordNet information is only informally represented (in • The key idea: exploit the rich information glosses and examples). --> Extended WordNet contained in the definitional glosses. • WordNet consists of different unrelated data-bases for common nouns, verbs, adjectives (and adverbs). - No information about • Intent is to automatically (1) syntactically parse cross-categorial sense distinctions. the glosses, (2) transform glosses into logical • Wide variation of granularity in different parts of WordNet. forms and (3) semantically tag the nouns, verbs, • In general, WordNet tends to be too fine-granular (branching factor adjectives and adverbs of the glosses. and depth of hierarchy). • Example: • WordNet focusses on paratactic semantic relations. No information on the level of predicate-argument structure. --> Thematic Roles – Excellent • No information about selectional constraints/ preferences. – Gloss: „of highest quality“ • WordNet (of course) does not solve the (notoriously hard) problem – Logical form: excellent(x1) ! of(x1,x2)& highest(x2)& of word-sense disambiguation. quality(x2) • http://xwn.hlt.utdallas.edu/ Semantic Theory, SS 2008 © M. Pinkal, S. Thater 13 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 14 More complex inversion-like Inverse Relations correspondences • The inverse relation (covered by WordNet‘s heterogeneous „antonymy“ relation) implicitly states a regularity about of the respective Mary gave Peter the book concepts involved to their arguments: Peter received the book from Mary taller_than(x,y) % shorter_than(y,x) John sold the car to Bill for 3,000! parent_of(x,y) % child_of(y,x) Bill bought the car from John for 3,000! like(x,y) % please(y,x) Semantic Theory, SS 2008 © M. Pinkal, S. Thater 15 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 16 Verb alternations Verb alternations The window broke John sells the book. A rock broke the window John broke the window with a rock The book sells for 19.95!. break3(x,y,z) |= break2(z,y) |= break1(y) Mary reads the book The plane flew to Frankfurt The book reads easy. John flew the plane to Frankfurt John flew Bill with the plane to Frankfurt fly3(x,y,z) |= fly2(z,y) |= fly1(y) Semantic Theory, SS 2008 © M. Pinkal, S. Thater 17 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 18 Predicate-argument structure correspondences Thematic Roles (Fillmore 1968) • Verbs with varying number of explicit • Thematic roles describe the conceptual argument positions, and varying realization participants in a situation in a generic way, independent from their grammatical realization. of "the same argument". • Thematic roles form a small, closed, and • (Quasi-)Equivalent sentences with different universally applicable inventory conceptual realization of "the same" semantic argument types. argument positions. • A typical role inventory might consit of the roles: Agent, Theme (Patient, Object),Recipient, Instrument, Source, Goal, Beneficiary, Experiencer Semantic Theory, SS 2008 © M. Pinkal, S. Thater 19 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 20 Role Annotation Examples RoleRole-linking linking, Informationexample • Linking information, provided in the lexicon, maps syntactic functions to semantic roles • An example: – [The window]pat broke give: SB ! Agent – [A rock ]inst broke [the window ]pat OA ! Theme – [John ]ag broke [the window ]pat [with a rock ]inst OD ! Recipient receive: SB ! Recipient – [Peter ]ag gave [Mary]rec [the book ]pat OA ! Theme – [Mary ]rec received [the book ]pat [from Peter ]ag OP-from ! Agent • Some linguistic theories try to model role linking by general principles (linking theory). No precise and complete linking theory is available. Semantic Theory, SS 2008 © M. Pinkal, S. Thater 21 Semantic Theory, SS 2008 © M. Pinkal, S. Thater 22 Thematic Roles The Role Dilemma • Allow to represent the semantic correspondence between (uses of) relational concepts in a • A closed inventory of 8 or 12 or even 20 roles systematic way – thereby supporting basic is not sufficient to describe the wealth of lexical-semantic inference. predicate-argument relations. • Support a systematic representation of the mapping between syntactic complements and Options: semantic argument positions (role-linking). • Use role names in a more or less arbitrary • Support the systematic description of selectional