The Global Wordnet Grid: anchoring languages to universal meaning Piek Vossen Irion Technologies/Free University of Amsterdam and Christiane Fellbaum Princeton.

  • Published on
    27-Mar-2015

  • View
    212

  • Download
    0

Transcript

Slide 1The Global Wordnet Grid: anchoring languages to universal meaning Piek Vossen Irion Technologies/Free University of Amsterdam and Christiane Fellbaum Princeton University Slide 2 Overview Wordnet, EuroWordNet background Architecture of the Global Wordnet Grid Mapping wordnets to the Grid Kyoto: an implementation of the Grid Slide 3 WordNet1.5 Developed at Princeton by George Miller and his team as a model of the mental lexicon. Semantic network in which concepts are defined in terms of relations to other concepts. Structure: organized around the notion of synsets (sets of synonymous words) basic semantic relations between these synsets http://www.cogsci.princeton.edu/~wn/w3wn.html http://www.cogsci.princeton.edu/~wn/w3wn.html Slide 4 Structure of WordNet Slide 5 EuroWordNet The development of a multilingual database with wordnets for several European languages Funded by the European Commission, DG XIII, Luxembourg as projects LE2-4003 and LE4-8328 March 1996 - September 1999 2.5 Million EURO. http://www.hum.uva.nl/~ewnhttp://www.hum.uva.nl/~ewnhttp://www.hum.uva.nl/~ewn http://www.illc.uva.nl/EuroWordNet/finalresults- ewn.htmlhttp://www.illc.uva.nl/EuroWordNet/finalresults- ewn.html Slide 6 ENGLISH Car Train Vehicle Inter-Lingual-Index Transport Road Air Water DomainsTop Ontology Device Object TransportDevice English Words vehicle cartrain 1 2 4 33 Czech Words dopravn prostednk autovlak 2 1 French Words vhicule voituretrain 2 1 Estonian Words liiklusvahend autokillavoor 2 1 German Words Fahrzeug AutoZug 2 1 Spanish Words vehculo autotren 2 1 Italian Words veicolo autotreno 2 1 Dutch Words voertuig autotrein 2 1 EuroWordnet architecture Slide 7 EuroWordNet Wordnets are unique language-specific structures: different lexicalizations differences in synonymy and homonymy different relations between synsets same organizational principles: synset structure and same set of semantic relations. Language independent knowledge is assigned to the ILI and can thus be shared for all language linked to the ILI: both an ontology and domain hierarchy Slide 8 Autonomous & Language-Specific voorwerp {object} lepel {spoon} werktuig{tool} tas {bag} bak {box} blok {block} lichaam {body} Wordnet1.5Dutch Wordnet bag spoon box object natural object (an object occurring naturally) artifact, artefact (a man-made object) instrumentality blockbody container device implement tool instrument Slide 9 Differences in structure Artificial Classes versus Lexicalized Classes: instrumentality; natural object Lexicalization differences of classes: container and artifact (object) are not lexicalized in Dutch artifact substance (kunststof) is lexicaled in Dutch not in English Should we include all lexicalized classes from all (8) languages? What is the purpose of different hierarchies? Slide 10 Artificial ontology: better control or performance, or a more compact and coherent structure. introduce artificial levels for concepts which are not lexicalized in a language (e.g. instrumentality, hand tool), neglect levels which are lexicalized but not relevant for the purpose of the ontology (e.g. tableware, silverware, merchandise ). What properties can we infer for spoons? spoon -> container; artifact; hand tool; object; made of metal or plastic; for eating, pouring or cooking Linguistic versus Artificial Ontologies Slide 11 Linguistic ontology: Exactly reflects the relations between all the lexicalized words and expressions in a language. Captures valuable information about the lexical capacity of languages: what is the available fund of words and expressions in a language. What words can be used to name spoons? spoon -> object, tableware, silverware, merchandise, cutlery, Linguistic versus Artificial Ontologies Slide 12 Wordnets versus ontologies Wordnets: autonomous language-specific lexicalization patterns in a relational network. Usage: to predict substitution in text for information retrieval, text generation, machine translation, word- sense-disambiguation. Ontologies: data structure with formally defined concepts. Usage: making semantic inferences. Slide 13 Inter-Lingual-Index: unstructured fund of concepts to provide an efficient mapping across the languages; Index-records are mainly based on WordNet synsets and consist of synonyms, glosses and source references; Various types of complex equivalence relations are distinguished; Equivalence relations from synsets to index records: not on a word-to-word basis; Indirect matching of synsets linked to the same index items; The Multilingual Design Slide 14 Equivalent Near Synonym 1. Multiple Targets (1:many) Dutch wordnet: schoonmaken (to clean) matches with 4 senses of clean in WordNet1.5: make clean by removing dirt, filth, or unwanted substances from remove unwanted substances from, such as feathers or pits, as of chickens or fruit remove in making clean; "Clean the spots off the rug" remove unwanted substances from - (as in chemistry) 2. Multiple Sources (many:1) Dutch wordnet: versiersel near_synonym versiering ILI-Record:decoration. 3. Multiple Targets and Sources (many:many) Dutch wordnet: toestel near_synonym apparaat ILI-records:machine; device; apparatus; tool Slide 15 Equivalent Hyperonymy Typically used for gaps in English WordNet: genuine, cultural gaps for things not known in English culture: Dutch: klunen, to walk on skates over land from one frozen water to the other Dutch: citroenjenever, which is a kind of gin made out of lemon skin, pragmatic, in the sense that the concept is known but is not expressed by a single lexicalized form in English: Dutch: kunstproduct = artifact substance artifact object Dutch: hoofd = human head and Dutch: kop = animal head, English uses head for both. Slide 16 From EuroWordNet to Global WordNet Currently, wordnets exist for more than 40 languages, including: Arabic, Bantu, Basque, Chinese, Bulgarian, Estonian, Hebrew, Icelandic, Japanese, Kannada, Korean, Latvian, Nepali, Persian, Romanian, Sanskrit, Tamil, Thai, Turkish, Zulu... Many languages are genetically and typologically unrelated http://www.globalwordnet.org Slide 17 Some downsides Construction is not done uniformly Coverage differs Not all wordnets can communicate with one another Proprietary rights restrict free access and usage A lot of semantics is duplicated Complex and obscure equivalence relations due to linguistic differences between English and other languages Slide 18 Inter-Lingual Ontology Device Object TransportDevice English Words vehicle cartrain 1 2 33 Czech Words dopravn prostednk autovlak 2 1 French Words vhicule voituretrain 2 1 Estonian Words liiklusvahend autokillavoor 2 1 German Words Fahrzeug AutoZug 2 1 Spanish Words vehculo autotren 2 1 Italian Words veicolo autotreno 2 1 Dutch Words voertuig autotrein 2 1 Next step: Global WordNet Grid Slide 19 GWNG: Main Features Construct separate wordnets for each Grid language Contributors from each language encode the same core set of concepts plus culture/language-specific ones Synsets (concepts) can be mapped crosslinguistically via an ontology No license constraints, freely available Slide 20 The Ontology: Main Features Formal, artificial ontology serves as universal index of concepts List of concepts is not just based on the lexicon of a particular language (unlike in EuroWordNet) but uses ontological observations Concepts are related in a type hierarchy Concepts are defined with axioms Slide 21 The Ontology: Main Features In addition to high-level (primitive) concept ontology needs to express low-level concepts lexicalized in the Grid languages Additional concepts can be defined with expressions in Knowledge Interchange Format (KIF) based on first order predicate calculus and atomic element Slide 22 The Ontology: Main Features Minimal set of concepts (Reductionist view): to express equivalence across languages to support inferencing Ontology must be powerful enough to encode all concepts that are lexically expressed in any of the Grid languages Slide 23 The Ontology: Main Features Ontology need not and cannot provide a linguistic encoding for all concepts found in the Grid languages Lexicalization in a language is not sufficient to warrant inclusion in the ontology Lexicalization in all or many languages may be sufficient Ontological observations will be used to define the concepts in the ontology Slide 24 Ontological observations Identity criteria as used in OntoClean (Guarino & Welty 2002), : rigidity: to what extent are properties true for entities in all worlds? You are always a human, but you can be a student for a short while. essence: what properties are essential for an entity? Shape is essential for a statue but not for the clay it is made of. unicity: what represents a whole and what entities are parts of these wholes? An ocean is a whole but the water it contains is not. Slide 25 Type-role distinction Current WordNet treatment: (1) a husky is a kind of dog(type) (2) a husky is a kind of working dog (role) Whats wrong? (2) is defeasible, (1) is not: *This husky is not a dog This husky is not a working dog Other roles: watchdog, sheepdog, herding dog, lapdog, etc. Slide 26 Ontology and lexicon Hierarchy of disjunct types: Canine PoodleDog; NewfoundlandDog; GermanShepherdDog; Husky Lexicon: NAMES for TYPES: {poodle}EN, {poedel}NL, {pudoru}JP ((instance x Poodle) LABELS for ROLES: {watchdog}EN, {waakhond}NL, {banken}JP ((instance x Canine) and (role x GuardingProcess)) Slide 27 Ontology and lexicon Hierarchy of disjunct types: River; Clay; etc Lexicon: NAMES for TYPES: {river}EN, {rivier, stroom}NL ((instance x River) LABELS for dependent concepts: {rivierwater}NL (water from a river => water is not Unit) ((instance x water) and (instance y River) and (portion x y) {kleibrok}NL (irregularly shared piece of clay=>Non-essential) ((instance x Object) and (instance y Clay) and (portion x y) and (shape X Irregular)) Slide 28 Rigidity The primitive concepts represented in the ontology are rigid types Entities with non-rigid properties will be represented with KIF statements But: ontology may include some universal, core concepts referring to roles like father, mother Slide 29 Properties of the Ontology Minimal: terms are distinguished by essential properties only Comprehensive: includes all distinct concepts types of all Grid languages Allows definitions via KIF of all lexemes that express non-rigid, non-essential properties of types Logically valid, allows inferencing Slide 30 Mapping Grid Languages onto the Ontology Explicit and precise equivalence relations among synsets in different languages, which is somehow easier: type hierarchy is minimal subtle differences can be encoded in KIF expressions Grid database contains wordnets with synsets that label either primitive types in the hierarchies, or words relating to these types in ways made explicit in KIF expressions If 2 lgs. create the same KIF expression, this is a statement of equivalence! Slide 31 How to construct the GWNG Take an existing ontology as starting point; Use English WordNet to maximize the number of disjunct types in the ontology; Link English WordNet synsets as names to the disjunct types; Provide KIF expressions for all other English words and synsets Slide 32 How to construct the GWNG Copy the relation from the English Wordnet to the ontology to other languages, including KIF statements built for English Revise KIF statements to make the mapping more precise Map all words and synsets that are and cannot be mapped to English WordNet to the ontology: propose extensions to the type hierarchy create KIF expressions for all non-rigid concepts Slide 33 Initial Ontology: SUMO (Niles and Pease) SUMO = Suggested Upper Merged Ontology --consistent with good ontological practice --fully mapped to WordNet(s): 1000 equivalence mappings, the rest through subsumption --freely and publicly available --allows data interoperability --allows NLP --allows reasoning/inferencing Slide 34 SUMO 1,000 generic, abstract, high-level terms 4,000 definitional statements MILO (Mid-Level Ontology) closer to lexicon, WordNet Slide 35 Mapping Grid languages onto the Ontology Check existing SUMO mappings to Princeton WordNet -> extend the ontology with rigid types for specific concepts Extend it to many other WordNet synsets Observe OntoClean principles! (Synsets referring to non-rigid, non-essential, non- unicitous concepts must be expressed in KIF) Slide 36 Lexicalizations not mapped to WordNet Not added to the type hierarchy: {straathond}NL (a dog that lives in the streets) ((instance x Canine) and (habitat x Street)) Added to the type hierarchy: {klunen}NL (to walk on skates from one frozen body to the next over land) KluunProcess => WalkProcess Axioms: (and (instance x Human) (instance y Walk) (instance z Skates) (wear x z) (instance s1 Skate) (instance s2 Skate) (before s1 y) (before y s2) etc National dishes, customs, games,.... Slide 37 Most mismatching concepts are not new types Refer to sets of types in specific circumstances or to concept that are dependent on these types, next to {rivierwater}NL there are many others: {theewater}NL (water used for making tea) {koffiewater}NL (water used for making coffee) {bluswater}NL (water used for making extinguishing file) Relate to linguistic phenomena: gender, perspective, aspect, diminutives, politeness, pejoratives, part-of-speech constraints Slide 38 {teacher}EN ((instance x Human) and (agent x TeachingProcess)) {Lehrer}DE ((instance x Man) and (agent x TeachingProcess)) {Lehrerin}DE ((instance x Woman) and (agent x TeachingProcess)) KIF expression for gender marking Slide 39 KIF expression for perspective sell: subj(x), direct obj(z),indirect obj(y) versus buy: subj(y), direct obj(z),indirect obj(x) (and (instance x Human)(instance y Human) (instance z Entity) (instance e FinancialTransaction) (source x e) (destination y e) (patient e) The same process but a different perspective by subject and object realization: marry in Russian two verbs, apprendre in French can mean teach and learn Slide 40 Part-of-speech mismatches {bankdrukken-V}NL vs.{bench press-N}EN {gehuil-N}NL vs. {cry-V}EN {afsluiting-N}NL vs. {close-V}EN Process in the ontology is neutral with respect to POS! Slide 41 Parallel Noun and Verb hierarchy event act deed sail promise change movement change of location to happen to act to do to sell a promise to change to move to move position Encoded once as a Process in the ontology! Slide 42 Mixed Noun and Adjective hierarchy Colour: red, blue, green, etc. Height: high, low Size: big, small Emotion: sad, angry, happy, anxious etc. Encoded once as a attributes in the ontology! Slide 43 Aspectual variants Slavic languages: two members of a verb pair for an ongoing event and a completed event. English: can mark perfectivity with particles, as in the phrasal verbs eat up and read through. Romance languages: mark aspect by verb conjugations on the same verb. Dutch, verbs with marked aspect can be created by prefixing a verb with door: doorademen, dooreten, doorfietsen, doorlezen, doorpraten (continue to breathe/eat/bike/read/talk). These verbs are restrictions on phases of the same process Which does NOT warrant the extension of the ontology with separate processes for each aspectual variant Slide 44 Aspectual lexicalization Regular compositional verb structures: doorademen: (lit. through+breath, continue to breath) doorbetalen:(lit. through+pay, continue to pay) doorlopen:(lit. through+walk, continue to walk) doorfietsen: (lit. through+walk, continue to walk) doorrijden: (lit. through+walk, continue to walk) (and (instance x BreathProcess)(instance y Time) (instance z Time) (end x z) (expected (end x y) (after z y)) Slide 45 MORE GENERAL VERBS: openmaken: (lit. open+make, to cause to be open); dichtmaken: (lit. close+make, to cause to be open); MORE SPECIFIC VERBS: openknijpen(lit. open+squeeze, to open by squeezing) has_hyperonymknijpen (squeeze) & openmaken (to open) opendraaien(lit. open+turn, to open by turning) has_hyperonymdraaien (to turn) & openmaken (to open) dichtknijpen: (lit. closed+squeeze, to close by squeezing) has_hyperonymknijpen (squeeze) & dichtmaken (to close) dichtdraaien: (lit. closed +turn, to close by turning) has_hyperonymdraaien (to turn) & dichtmaken (to close) Lexicalization of Resultatives Slide 46 Kinship relations in Arabic (Eam~)father's brother, paternal uncle. (xaAl)mother's brother, maternal uncle. (Eam~ap)father's sister, paternal aunt. (xaAlap)mother's sister, maternal aunt Slide 47 Kinship relations in Arabic......... ($aqiyqapfull) sister, sister on the paternal and maternal side (as distinct from (>uxot): 'sister' which may refer to a 'sister' from paternal or maternal side, or both sides). (vakolAna)father bereaved of a child (as opposed to (yatiym) or (yatiymap) for feminine: 'orphan' a person whose father or mother died or both father and mother died). (vakolaYa)other bereaved of a child (as opposed to or for feminine: 'orphan' a person whose father or mother died or both father and mother died). Slide 48 father's brother, paternal uncle WORDNET paternal uncle => uncle => brother of....???? ONTOLOGY (=> (paternalUncle ?P ?UNC) (exists (?F) (and (father ?P ?F) (brother ?F ?UNC)))) Complex Kinship concepts Slide 49 Fine tune equivalence relations {rivier}NL (and (instance x River) (instance y RiverMouth) (instance z Country) (part y x) (location y z) {stroom}NL (and (instance x River) (instance y RiverMouth) (instance p RiverPart) (not (equal p y) (instance z Country) (location p z) (not (location y z)) Slide 50 Universality as evidence If lexicalization of the specific process is more universal it can be seen as evidence that the specific processes should be listed in the ontology and not the generic verb: English verb cut abstracts from the precise process but there are troponyms that implicate the manner : snip, clip imply scissors, chop and hack a large knife or an axe Dutch there is no general verb but only specific verbs: knippen clip, snip, cut with scissors or a scissor-like tool', snijden cut with a knife or knife-like tool, hakken chop, hack, to cut with an axe, or similar tool). If Father is lexicalized in most languages we add it to the ontology even when it is NOT Rigid! Slide 51 Universality as evidence Artifact substance is lexicalized in Dutch and other languages => ArtifactObject in SUMO needs to be generalized to Artifact so that it can be applied to both substances and objects Slide 52 Open Questions/Challenges What is a word, i.e., a lexical unit? What is the status of complex lexemes like English lightning rod, word of mouth, find out, kick the bucket? What is the status of compounds in Germanic languages and Chinese? "hottentottententententoonstelling" (exposition of tents of the "hottentotten" (African tribe)) What is a semantic unit, i.e. a concept? Slide 53 Availability "buiten dienst" = out of service "buitendienst" = peripheral service "hottentottententententoonstelling" = exposition of tents of the "hottentotten" (African tribe) Slide 54 Open Questions/Challenges Is there a core inventory of concepts that are universally encoded? If so, what are these concepts? How can crosslinguistic equivalence be verified? Is there systematicity to the language-specific extensions? What are the lexicalization patterns of individual languages? Are lexical gaps accidental or systematic? Slide 55 Coverage: what belongs in a universal lexical database? Formal, linguistic criteria for inclusion Informal, cultural criteria Both are difficult to define and apply! Slide 56 Concrete goals for GWG Global Wordnet Association website: http://www.globalwordnet.org/gwa/gwa_grid.htm 5000 Base Concepts or more: English Spanish Catalan Czech, Polish, Dutch, other wordnets 7 th Frame Work project Kyoto Slide 57 KYOTO Project 7 th Frame Work project (under negotiation) Kowledge Yielding Ontologies for Transition-based Organisations Goal: Global Wordnet Grid = ontology + wordnets AutoCons = Automatic concept extractors Kybots = Knowledge yielding robots Wiki environment for encoding domain knowledge in expert groups Index and retrieval software for deep semantic search Languages: Dutch, English, Spanish, Basque, Italian, Chinese and Japanese Domain of application: environmental organisations Period: March/April 2008 - 2011 Slide 58 KYOTO Consortium Universities Vrije Universiteit Amterdam, Amsterdam, Netherlands Consiglio Nazionale delle Ricerche, Pisa, Italy Berlin-Brandenburg Academy of Sciences and Humantities, Berlin, Germany Euskal Herriko Unibertsitatea, San Sebastian, Spain Academia Sinica, Taipei, Taiwan National Institute of Information and Communications Technology, Kyoto, Japan Masaryk University, Brno, Czech Companies Irion Technologies, Delft, Netherlands Synthema, Pisa, Italy Users European Centre for Nature Conservation, Tilburg, Netherlands World Wide Fund for Nature, Zeist, Netherlands Slide 59 Environmental organizations Capture Index Docs URLs Experts Images Search Dialogue Concept Mining Fact Mining AbstractPhysical Top Middle Domain waterCO2 Substance CO2 emission water pollution Universal Ontology Wordnets Environmental organizations Citizens Governors Companies Domain Wiki Process Slide 60 Text & Meta data in XMLFormat term hierarchy wordnet Concept Miners term relations ontology Kybots Manual Revision Wiki DEB Client 2 3 5 domain wordnet domain ontology Indexing source data Capture Data & Facts in XML Format DEB Server Access end-users Index 6 User scenarios User scenarios Manual Test Bench mark data Bench marking 1 1 4 7 8 Slide 61 AbstractPhysical waterCO2 Substance CO2 emission water pollution OntologyWordnets Generic Process Chemical Reaction Logical ExpressionsLinguistic Miners or Kybots Domain words Slide 62 END

Recommended

View more >