Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Situations as attractors for semantic interpretation

423 views

Published on

slides used for the presentation at DGFS 2006

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Situations as attractors for semantic interpretation

  1. 1. Situations are“Attractors” of Semantic Interpretations Kow KURODA* Keiko NAKAMOTO** Hajime NOZAWA* Hitoshi ISAHARA**National Institute of Information and Communications Technology (NICT), Japan **Department of Education, Kyoto University, Japan presented at Corpus-based Approaches to Noncompositional Phenomena, A DGfS Workshop Feb 23, 2006Note: This PDF is a version with some modifications after the presentation
  2. 2. Overview• Brief explanation for our previous work (Nakamoto, Kuroda and Nozawa 2005), an exhaustive corpus- based detailed sematic analysis of Japanese verb osou inspired by the Frame Semantics/FrameNet approach (Fillmore 1985; Fillmore and Atkins 1994; Fillmore, et al. 2003)• Define the Attraction-to-Situation Hypothesis • Propose a background theory of meaning construction • Define predictions to test experimentally based on the proposed theory• Conclusion 2
  3. 3. Corpus Analysis
  4. 4. Background• A detailed semantic analysis of Japanese verb osou (413 instances in total) was done exhaustively against a corpus (500,000 Japanese-English alignments (JEA corpus; Utiyama and Isahara 2003)), adopting the Frame Semantics/FrameNet approach• Main Results • Uses of osou are covered by 15 “situations,” or “semantic frames” at reasonably finer-grained granularities, including metaphoric and metonymic uses. • Figurative uses are bounded in that specific situations like <Bank robbery>, <Invasion>, <Suffering a Disaster> that serve as “attractors” of semantic interpretations. 4
  5. 5. A previous corpus- Semantic 112 L0 = Sub Semantic Classes at Semantic Classes English verbs that translate OSOU L1 Classes at L2 L3 (TOTAL) L1 Level Level 2 at Level 3 Level 1 Resource-based analysis revealed Intended Harm- attack[+human(s)]: rob 4 7 10 threatenig 51 90 Cause oriented causation[+animate] situations attack[+human(s)]: rob: break into 2 attack[+human(s)]: rob: make off with MONEY 1the usage, both literal attack[+human(s)]: rob: hold up attack[+human(s)]: rob: threaten 1 2 3 Life-and figurative, of attack[+human(s)] 23 23 42 threatening by human attack[+human(s)]: kill 1 1 attack[+human(s)]: assault 9 10Japanese verb osou (and attack[+human(s)]: assault: raid attack[+human(s)]: assault: shoot attack[+human(s)]: assault: shoot, wound 1 3 1 5its English counterparts) attack[+human(s)]: assault: shoot; rob 1 attack[+human(s)]: assault: stab 3 3 Life- attack[-human(s),+animal(s)] 7 8 9 threatening bycan be described for in attack[-human(s),+animal(s)]: kill attack[-human(s),?animal]: assault[+metaphoric?]: 1 1 1 nonhumanterms of situation/frame turn on Natural Disasters = Harm- hit,strike: hit 3 8 18 39 disasters causation[-animate] hit,strike: rock 1hierarchy (in next slide) hit,strike: strike 2 hit,strike: pound 2 hit,strike: destroy: wreak on 1 2 hit,strike: destroy: ravage 1comprising of roughly hit,strike: roar through hit,strike: sweep through hit,strike: wrought devastation 1 1 1 2 615 situations (F01-F15) hit,strike: IMPLICIT in: earthquake 2 hit,strike: IMPLICIT in: in PLACE 2 hit,strike: there is 1 Socialat the lowest, most hit,strike[+metaphoric, +human(s)?]: occur[=attack] hit,strike[+metaphoric]: hurt 1 1 2 21 disasters[+met aphoric]specific levels. hit,strike[+metaphoric?]: hit 2 9 hit,strike[+metaphoric]: hit 5 hit,strike[+metaphoric]: paralyze 1 hit,strike[+metaphoric]: IMPLICIT in: shocks from 1 hit,strike[+metaphoric]: overtake 1 4Note: sampling in this table is partial: hit,strike[+metaphoric]: take a toll hit,strike[+metaphoric]: besiege 1 1 hit,strike[+metaphoric]: engulf 1112/413: JEA corpus has two hit,strike[+metaphoric]: occur 2 4 hit,strike[+metaphoric]: fall on 1components: public and protected. 112 hit,strike[+metaphoric]: IMPLICIT in: in PLACE hit,strike[+metaphoric]: IMPLICIT in: problems 1 1 2is the number of osou’s instances found hit,strike[+metaphoric]: IMPLICIT in: turmoil 1 Sufferings = Harm- suffer 3 5 10 Sufferings 10 10 Effect orientedin the public component. 413 is the suffer: IMPLICIT in: victim 1 experiencenumber of osou’s occurrences in the suffer: be injured suffer: feel pain 1 1 3entire JEA corpus. suffer: bring sorrow to people 1 suffer: feel anxiety 1 suffer: seized with 1 4 suffer: suddenly begin a SYMPTOM 1 5 suffer: experience attack[-human(s), +metaphoric] 2
  6. 6. More Abstract More Concrete 狼が羊の群れを襲った L2 Level Situations L1 Level Situations Wolves {attacked; ?*assaulted} a flock of sheep. F06: Predatory Victimization MM 1a F07b: (Counter)Attack for スズメバチの群れが人を襲った Self-defense A swarm of wasps {attacked; ?*assaulted} people. A: Victimization MM 1c of Animal by MM 1b Animal F07: サルの群れが別の群れを襲った Nonpredatory A group of apes {attacked; ?assaulted} another group. Victimization F07a: Territorial B2: Victimization Conflict between A,B: of Human by MM 1d Groups F01: Conflict マフィアの殺し屋が別の組織の組長を襲った Animal Victimization of between Human A hitman of a Mafia {attacked; assaulted} the leader of the Animal by MM* 2 Groups Animal E: Conflict MM 2 opponents. between Groups 暴徒と化した民衆が警官隊を襲った B: Victimization F01,02: Power of Human by Conflict between A mob {attacked; ?assaulted} the squad of police. Animal Human Groups MM 1e F01,02,03: 貧しい国が石油の豊富な国を襲った Resource-aiming F02: Invasion A poor country {attacked; ??assaulted} the oil-rich country. Victimization B1: Victimization B1a: Physical of Human by Hurting = F03: Robbery 三人組の男が銀行を襲った. Human Violence A gang of three {attacked; ??assaulted} the bank branch. F04: Persection 狂った男が小学生を襲った B1b: Abuse A crazy man {attacked, assaulted} boys at elementary school. F05: RapingA,B,C,D,E (=ROOT): 男が二人の女性を襲ったVictimization of Y by X A man {attacked; assaulted; ??hit} a young woman. 突風がその町を襲った F09: Natural Disaster ?MM 7b on Smaller Scale Gust of wind {?*attacked; hit; ?*seized} the town. F09,10(,11): Natural Disaster MM 7a MM 0 ?MM 6b F10: Natural Disaster 地震がその都市を襲った on Larger Scale ? An earthquake {*attacked; hit; ?*seized} the city. D: Perceptible Impact MM 6a F11: Epidemic Spead MM 3a ペストがその町を襲った The Black Death {?*attacked; hit; ?seized} the town. MM 3b F12a: Social Disaster on ? Larger Scale C,D,E: 大型の不況がその国を襲った Victimization by Disaster C: Disaster A big depression {?*attacked; hit; ???seized} the country. F12: Social Disaster MM 5a MM 4a F12b: Social Disaster on ? Smaller Scale 赤字がその会社を襲った MM 8 MM 5b The company {experienced; *suffered; went into} red figures. (cf. Red figures {?attacked; ?hit; ?*seized} the company}) ? 肺癌が彼を襲った Hierarchical Frame Network (HFN) F13: Long-term He {suffered; was hit by} a lung cancer sickness of “X-ga Y-wo osou” (active) and F13,14: Suffering a MM 4b (cf. Cancer {??attacked; hit; seized} him) “Y-ga X-ni osowareru” (passive) F13,14,15: Getting Sick = Suffering a Mental or Physical Disorder 痙攣が患者を襲った Physical Disorder The patient have a convulsive fit E: Personal F14: Short-term (cf. A convulsive fit {??attacked; ?seized him) Disaster? sickness F14,15: Temporal ?MM 4cNOTES Suffering a Mental or 無力感が彼を襲った Physical Disorder• Instantiation/inheritance relation is indicated by solid arrow. He {suffered from; was seized by} inertia• Typical “situations” at finer-grained levels are thick-lined. F15: Short-term mental (cf. The inertia {?*attacked; ?hit; ?seized} him).• Dashed arrows indicate that instantiation relations are not disorderguaranteed. L1 Level Situations 不安が彼を襲った ?• attack is used to denote instantiations of A, B. He was seized with a sudden anxiety.• assault is used to denote instantiations of B1. L2 Level Situations (cf. Anxiety attacked him suddenly}• hit, strike are used to denote instantiations of C.• Pink arrow with MM i indicates a metaphorical mapping: 暴走トラックが子供を襲ったSource situations are in orange. F08: Accident The children got victims of a runaway truck (cf. A runaway truck {*attacked; ?*hit} children.) 6
  7. 7. Correspondence to Frames in Berkeley FrameNet• Berkeley FrameNet provides three relevant frames: • <Attacking> • <Cause_impact> • <Cause_harm> • the first two of which correspond to pretty generic situations of <(Intended) Harm-causation by Animate>, and <Unintended Harm-causation by Inanimate>, i.e., two semantic classes at the L2 level granularity. respectively • It is not clear what situation <Cause_harm> frame corresponds to.
  8. 8. Measuring Metaphoricality• Loose Metaphoricality Index: Nonmetaphor = “Threatenings by Human or Nonhuman” • 0.455 = 1.0 - 62/ 112 [metaphors/all uses]• Strict Metaphoricality Index 1: Nonmetaphor = “Life- or Resource-Threatenings by Human” • 0.536 = 1.0 - 52/ 112 [metaphors/all uses]• Strict Metaphoricality Index 2: Nonmetaphor = “Life- threatenings by Human or Nonhuman” • 0.545 = 1.0 - 51/112 [metaphors/all uses] 8
  9. 9. Why Did We Do This?• To estimate how much finer-granularity would be needed if we decide on providing “realistic” semantic analysis/annotation as your goal • An on-going work on this will be presented at LREC 2006• (Even) FrameNet frames turned out to be (too) coarse-grained to give a realistic specification of the “understood content” in unrestricted fashion. • “realistic” means specifying what (average) people understand when they hear a sentence (wether in or out of context) as precise as is justifiable with experiments. • “unrestricted” means no assumed specific applications: Information Retrieval, QA, Machine Translation 9
  10. 10. Metaphoric Uses “Dominate”• Both “X-ga Y-wo osou” [active] and “Y-ga X-ni osowareru” [passive] were examined • While the two forms have different preferences for situations. • Virtually, more than half of the uses turned out to be metaphorical• But • simply knowing how to detect metaphorical uses of a word is not itself a “solution” to the most serious problem of specifying the “target” meaning of each meaning transfer/metaphor. 10
  11. 11. What Drives Metaphorical Uses: Targets or Sources?• Metaphorical uses would be more systematic than you expect, in that they look “bounded”.• But they have a sparse distribution and therefore are less systematic than can be accounted for by Conceptual Metaphor/Mapping Theory (Lakoff and Johnson 1980, 1999) • A set of conceptual metaphors like <Lust Is Hunger> just allow you to specify the necessary conditions for figurative uses; not the sufficient conditions: a lot of idiosyncratic — yet “understandable”— constraints are found for each metaphor• But not enough time to elaborate on details today 11
  12. 12. Why Do Metaphoric Uses Look Bounded? • Hard question to answer, but there is a hint: • What seems to determine the range of metaphoric uses is the “targets,” not “sources,” of metaphorical mappings/ meaning transfers. • This led us to the idea/hypothesis that • the semantic interpretation of sentence s = w1 w2 ... wn is “attracted” to a closed set of specific, understandable situations specifiable in terms of semantic frames at finer- grained levels. 12
  13. 13. And when such attraction takes place• (lexical) meanings {m1, m2, ..., mn} (mi = m(wi)) of the words, {w1, w2, ..., wn}, are “adjusted” to the meaning of s.• This is a Gestalt effect in that (the meanings of) a “whole” and its “parts” are given at the same time. • This is just a special case of “accommodation to a schema” in the Piagetian theory of conceptual development. • and it is an example of “semantic accommodation” in Cognitive Grammar framework (Langacker 1987, 1991, among others) • and it is also an example of “co-composition” in Generative Lexicon framework (Pustejovsky 1995)• But refinements are in need to get the basic ideas to work. 13
  14. 14. A “Competitive” Theory of Meaning Construction— Defining a Framework —
  15. 15. A Theory to Test [1/2]• Meaning construction is a competitive process like Darwinian natural selection, rather than (just) a (co)- compositional process, in that it has the “winner-take- all” property.• A sentence “evokes” a set of “candidate” interpretations, each with a “goodness of fit” measured against “meaningfulness” models M* = {M1, M2, ..., Mn}.• Interpretation with the highest goodness of fit score wins out (cf. Optimality Theory) 15
  16. 16. A Theory to Test [2/2]• Interpretive models M is a set of conceivable situations, or “semantic frames” at finer-granularities.• This allows the following: • Each word in a sentence S evokes a set of unrelated situation/frames independently. • Yet meaning construction for S is successful as far as situation-evocation “converges” after a “selection” process • (A lot of) semantics is “distributed” over combinations of words, rather than “symbolized” by single words. • as claimed in Collostructional Analysis (Stefanowitsch and Gries 2003; Gries and Stefanowitsch 2004) 16
  17. 17. How Competition Converges”Winner” (Sub)frames Frame[1] Frame Element[1]: ... • Each “semantic unit” SU, not Frame Element[2]: ... necessarily a word, within a ... activates activates Frame Element[n]: ... sentence “activates” a set of SU[1] Definition: ... activates (sub)frames independently. activates accomodates Frame[j] Frame[i] activates • Evoked (sub)frames compete SU[i] activates Frame Element[1]: ... each other either by Frame Element[2]: ... inhibits activates accomodates ... Frame Element[n]: ... • “activation” • SU[n] Definition: ... or “(lateral) inhibition” inhibits activates inhibits inhibits • Once competition settles inhibits Frame[k] down, the (meaning of) SUs of Frame Element[1]: ... Frame Element[2]: ... the “loser” (sub)frames ... Frame Element[n]: ... accommodate to the Definition: ...”Loser“ (Sub)frame(s) (meanings of) “winner” (sub)frames 17
  18. 18. Remarks• Some crucial aspects of the proposed model of meaning construction is not new at all. • The basic ideas come from Parallel Distributed Processing (PDP) model of cognitive processes (Rumelhart, et al. 1985; McClelland, et al. 1986) • A PDP-style situation/frame selection was already implemented in G. Cottrell’s handcrafted, hardwired neural network for Word Sense Disambiguation (Cottrell 1985)• But the proposed model still has something new to it. • Word senses are “modified” or even “generated” through their accommodation to the meaning of a sentence that they occur within. 18
  19. 19. Attraction-to-Situation Hypothesis• As a result of “selection for goodness of fit,” semantic interpretations of a given sentence S are “attracted” to (ideally) one of the possible and very likely situations.• This predicts the following: • Attraction-to-Situation Hypothesis: Attraction to situation is effective even if all arguments are not explicitly given.• This can be tested experimentally using “semantic feature rating” (SFR) method (Nakamoto, Kuroda and Nozawa 2005) • SFR: a word within a sentence is rated against a set of semantic/characteristic features 19
  20. 20. Prediction to Test• If the theory is correct, a nonce word w* is feature- rated very much like a real word, if its occurring context C(w*) = W(s) – w* evokes a situation strong enough.• For English examples, • C(w* for Victim) = ___ {was attacked; was hit; suffered (from)} <Harm-causer> • C(w* for Harm causer) = <Victim> {was attacked; was hit; suffered (from)} ___• Is this a right prediction? — Let’s test it! 20
  21. 21. Testing Procedure• Japanese sentence X-ga Y-wo osou [active], or Y-ga X-ni osowareru [passive]) ALWAYS denotes a situation in which Victimization of Y by a Harm-causer X occurs.• Prediction confirms if SFRs for nonce words in C1, C2 conditions are similar to real words in C0 • C0: {X: Real, Y: Real, osou} neutral • C1: {X: Real, Y: Nonce, osou} attraction by Harm-causer • C2: {X: Nonce, Y: Real, osou} attraction by Victim 21
  22. 22. Subject NP PP denoting Harm- Transliteration (word-by-wordSituation ID Translation Original Example (in Passive form) denoting Victim causer translation from Japanese)F01: Power Conflict The President was attacked by an The President was assaulted by an The President an assasin !"#$% &( )*+,-between Human Groups assasin. assasin. an(other) armed A country was attacked by another A country was attacked by anotherF02: Invasion a country ./0% 12345( )*+,- country armed country. armed country.F03: Robbery on larger A bank branch was attacked by a A bank branch was attacked by a a bank branch a masked man 678% 9:;<=>( )*+,-scales masked man. masked man.F03: Robbery on smaller An old lady was attacked by a purse An old lady was assaulted by a purse an old lady a purse snatcher !"?8% @A( )*+,-scales snatcher. snatcher. Passengers-by were attacked by a Passengers-by were assaulted by aF04: Persecution,Violence passengers-by a lunatic man !"BC% DEFGH( )*+,- lunatic man. lunatic man. A woman was sexually assaulted aF05: Rape, Sexually assault a woman a pervert A woman was attacked by a pervert. !"I% JKI( )*+,- pervert.F06: Preying animal attack; zebras lions Zebras were attacked by lions. Zebras were attacked by lions. !"LM% NOPQRS( )*+,-PredationF07: Nonpreying animal !"TU% VWXYZW[=7( ) children wasps Children were attacked by wasps. Children were attacked by wasps.attack, usually for defence *+,- A family was attacked by a runaway A family got the victim of an accident byF08: Accident a family a unrunaway truck ]^U_% ]`=Na( )*+,- truck. a runaway truck.F09: Natural disater on a town gust of wind A town was attacked by gust of wind. A town was hit by gust of wind. bcdc% QVeW( )*+,-smaller scalesF10: Natural disaster on an area a hurricane An area was attacked by a hurricane. An area was hit by a hurricane. f7% gh=>( )*+,-larger scalesF11: Epidemic spread a city influenza A city was attacked by influenza. A city was hit by influenza. !"?i% jk,lm( )*+,- The stock market was attacked by aF12: Social disaster the stock market a debacle The stock market was hit by a sharp fall. !">n% Do( )*+,- sharp fall. A man was seized by cancer; A manF13: Long-term sickness a man cancer A man was attacked by cancer. !"pa% q( )*+,- suffered cancer.F14,15: Short-term mental an old man panic An old man was attacked by panic. An old man was seized by panic. !"rn% nstuC( )*+,-disorder OR sickness A man was seized by a sharp pain; A manF14: Short-term sickness a man a sharp pain A man was attacked by a sharp pain. vwx% yzC( )*+,- suffered a sharp pain.F15: Short-term mental A young man was attacked by a strong A young man was seized by a strong a young man strong jealousy !"8% {n=|W( )*+,-disorder jealousy jealousyNONE zebras a masked man Zebras were attacked by a masked man. Zebras were attacked by a masked man. bcdc% gh=>( )*+,- Experiments
  23. 23. 24 features used for SFRClass ID English translation of rated feature Rated feature in Japanese Harm-causer 1 X is a living thing. X!"#$%&( Harm-causer 2 X chose Y for its target. X!Y)*+%!,- Harm-causer 3 X is visible. X!".#/01%&2 Harm-causer 4 X couldnt help doing it to Y. X3Y)!,-4!564789:;,-( Harm-causer 5 X is human. X!<=%&( Harm-causer 6 X had an aim to do so. X!">)?,@!,-( Harm-causer 7 X is a natural phenomenon. X!A$BC%&( Harm-causer 8 X did so to satisfy its desire or needs. X!AD4EF)G-H-I.!,-( Harm-causer 9 X planned to take off something from Y. X!YJKLJ)MNOPQ;,-( Harm-causer 10 X is the name for a sickness. X!RS%&( Harm-causer 11 Xs activity can kill Y. X!Y)!,@T7U9:3&( Harm-causer 12 X is a collection of living things. X!"#$4VWQ%&( Victim 1 Y is a living thing. Y!"#$%&( Victim 2 Y had a good chance to prepare for Xs activity. Y3X4!X.Y/4!Z[;,-( Victim 3 Y had some reason to be victimized by X. Y.!X.!]LKJ4^_3&,-( Victim 4 Y is human. Y!<=%&( Victim 5 Y was aware of being victimized by X. Y!X.!]`ab.cd8@8-( Victim 6 Xs activity on X may cause X to die. Y!X.!]-43ef%Tg9:3&( Victim 7 Y is the name for a place. Y!hi)jHSk%&( Victim 8 Y could avoid Xs harm. Y!X.!]4)lmH9:P%#-( The degree of Ys affectedness is greater than the Victim 9 Y4!]6!n</n14%o)p/( individual scale. Victim 10 Y suffered a harm by Xs activity on Y. q!X.!]@0r.st)uv-( Victim 11 Y has been targeted by X long before. q!X.wkJKx]@8-( Victim 12 Y itself might have invited Xs activity on it. Y3X.!]-4.!Y.Pyz3&( 23
  24. 24. Full sentences (Baseline) Transliteration (word-by-word Subject NP PP denoting Harm-Situation ID translation from Japanese assuming that More Natural Translation Original Example (in Passive form) denoting Victim causer osou translates to attack)F01: Power Conflict The President was attacked by an The President was assaulted by an The President an assassin !"#$ %&( )*+,-between Human Groups assassin. assasin. an(other) armed A country was attacked by another A country was attacked by anotherF02: Invasion a country ./0$ 120( )*+,- country armed country. armed country.F03: Robbery on larger A bank branch was attacked by a A bank branch was attacked by a masked a bank branch a masked man 34$ 5678( )*+,-scales masked man. man.F03: Robbery on smaller An old lady was attacked by a purse An old lady was assaulted by a purse an old lady a purse snatcher ./9:$ ;<,=>( )*+,-scales snatcher. snatcher. Passengers-by were attacked by a lunatic Passengers-by were assaulted by a lunaticF04: Persecution,Violence passengers-by a lunatic man ?4@$ ABCD78( )*+,- man. man. A woman was sexually assaulted aF05: Rape, Sexual assault a woman a pervert A woman was attacked by a pervert. ./EF$ FGHI( )*+,- pervert.F06: Preying animal attack; zebras lions Zebras were attacked by lions. Zebras were attacked by lions. JKLK$ MNOP( )*+,-PredationF07: Nonpreying animal children wasps Children were attacked by wasps. Children were attacked by wasps. QRS$ TUVWX( )*+,-attack, usually for defence A family was attacked by a runaway A family got the victim of an accident byF08: Accident a family a unrunaway truck ./YZ$ []M^_( )*+,- truck. a runaway truck.F09: Natural disater on a town gust of wind A town was attacked by gust of wind. A town was hit by gust of wind. ./`a$ bc( )*+,-smaller scalesF10: Natural disaster on an area a hurricane An area was attacked by a hurricane. An area was hit by a hurricane. ./de$ fc( )*+,-larger scales ./gh$ NPijkPl7m4( )F11: Epidemic spread a city influenza A city was attacked by influenza. A city was hit by influenza. *+,- a debacle (or a The stock market was attacked by aF12: Social disaster the stock market The stock market was hit by a sharp fall. nohp$ nq7[a( )*+,- downturn, sharp fall) sharp fall. A man was seized by cancer; A manF13: Long-term sickness a man cancer A man was attacked by cancer. ./@$ rF7sP( )*+,- suffered cancer.F14,15: Short-term mental an old man panic An old man was attacked by panic. An old man was seized by panic. ./9@$ tu( )*+,-disorder OR sickness A man was seized by a sharp pain; A manF14: Short-term sickness a man a sharp pain A man was attacked by a sharp pain. ./8F$ vw( )*+,- suffered a sharp pain.F15: Short-term mental A young man was attacked by a strong A young man was seized by a strong a young man strong jealousy ./x$ vyz{|( )*+,-disorder jealousy jealousyNONSENSICAL zebras a masked man Zebras were attacked by a masked man. Zebras were attacked by a masked man. JKLK$ 5678( )*+,- 24
  25. 25. Test sentences with a nonce word for Victim Subject NP PP denoting Harm- Transliteration (word-by-word translationSituation ID Translation denoting Victim causer from Japanese)F01: Power Conflict between Nonce Word an assasin ____ was attacked by an assasin. The President was assaulted by ___.Human Groups an(other) armed ____ was attacked by another armedF02: Invasion Nonce Word A country was attacked by ___. country country.F03: Robbery on larger scales Nonce Word a purse snatcher ____ was attacked by a purse snatcher. An old lady was assaulted by ___.F03: Robbery on smaller scales Nonce Word a masked man ____ was attacked by a masked man. A bank branch was attacked by ___.F04: Persecution,Violence Nonce Word a lunatic man ____ were attacked by a lunatic man. Passengers-by were assaulted by ___.F05: Rape; Sexually assault Nonce Word a pervert ____ was attacked by a pervert. A woman was sexually assaulted by ___.F06: Preying animal attack; Nonce Word lions ____ were attacked by lions. Zebras were attacked by ___.PredationF07: Nonpreying animal attack, Nonce Word wasps ____ were attacked by wasps. Children were attacked by ___.usually for defence A family got the victim of an accident byF08: Accident Nonce Word a unrunaway truck ____ was attacked by a runaway truck. ___.F09: Natural disater on smaller Nonce Word gust of wind ____ was attacked by gust of wind. A town was hit by ___.scalesF10: Natural disaster on larger Nonce Word a hurricane ____ was attacked by a hurricane. An area was hit by ___.scalesF11: Epidemic spread Nonce Word influenza ____ was attacked by influenza. A city was hit by ___.F12: Social disaster Nonce Word a debacle ____ was attacked by a sharp fall. The stock market was hit by ___. A man was seized by ___; A manF13: Long-term sickness Nonce Word cancer ____ was attacked by cancer. suffered (from) ___.F14,15: Short-term mental A man was seized by ___; A man Nonce Word a sharp pain ____ was attacked by a sharp pain.disorder OR sickness suffered (from) ___.F14: Short-term sickness Nonce Word panic ____ was attacked by panic. An old man was seized by ___.F15: Short-term mental Nonce Word strong jealousy ____ was attacked by a strong jealousy A young man was seized by ___.disorderNONE Nonce Word Nonce Word ____ were attacked by ___. ___ were attacked by ___. 25
  26. 26. Test sentences with a nonce word for Harm-causer Subject NP PP denoting Harm- Transliteration (word-by-word translationSituation ID Translation denoting Victim causer from Japanese)F01: Power Conflict between The President Nonce Word The President was attacked by ___. ____ was assaulted by an assasin.Human Groups ____ was attacked by another armedF02: Invasion a country Nonce Word A country was attacked by ___. country.F03: Robbery on larger scales an old lady Nonce Word An old lady was attacked by ___. ____ was assaulted by a purse snatcher.F03: Robbery on smaller a bank branch Nonce Word A bank branch was attacked by ___. ____ was attacked by a masked man.scalesF04: Persecution,Violence passengers-by Nonce Word Passengers-by were attacked by ___. ____ were assaulted by a lunatic man.F05: Rape, Sexually assault a woman Nonce Word A woman was attacked by ___. ____ was sexually assaulted by a pervert.F06: Preying animal attack; zebras Nonce Word Zebras were attacked by ___. ____ were attacked by lions.PredationF07: Nonpreying animal attack, children Nonce Word Children were attacked by ___. ____ were attacked by wasps.usually for defence ____ got victimized of an accident by aF08: Accident a family Nonce Word A family was attacked by ___. runaway truck.F09: Natural disater on a town Nonce Word A town was attacked by ___. ____ was hit by gust of wind.smaller scalesF10: Natural disaster on larger an area Nonce Word An area was attacked by ___. ____ was hit by a hurricane.scalesF11: Epidemic spread a city Nonce Word A city was attacked by ___. ____ was hit by influenza.F12: Social disaster the stock market Nonce Word The stock market was attacked by ___. ____ was hit by a debacle. ____ was seized by cancer; ___ sufferedF13: Long-term sickness a man Nonce Word A man was attacked by ___. (from) cancer.F14,15: Short-term mental ____ was seized by a sharp pain; ___ a man Nonce Word A man was attacked by ___.disorder OR sickness suffered a sharp pain.F14: Short-term sickness an old man Nonce Word An old man was attacked by ___. ____ man was seized by panic.F15: Short-term mental a young man Nonce Word A young man was attacked by ___. ____ man was seized by a strong jealousy.disorderNONE Nonce Nonce Word ___ were attacked by ___. ____ were attacked by ___. 26
  27. 27. $E* ,EPFKFFJ ,EP6HILJ ,EPE**H6 )EPFE66I ,EP*LILG )EPF*6*H ,EP*FHK6$EF )EPEGII6 ,EPGIGG6 )EPFJ6FF ,EPEG6IL )EP66KF* )EPE6JKG )EP*EJKG$EG )EPFIKHK ,EP6KHHE ,EPEJEGI )EPEEFIF ,EPFGF*J ,EP**F*H ,EPJHJEI$EH ,EP*HGK* ,EPFGJ*I )EPEG66G )EPHGIIL )EP6IIJ* ,EP6GEEK ,EPEJFJF$EI ,EPEFFH* ,EP**IL* ,EPI*6II )EP*L*G* )EPEKHFH ,EPFKE*H )EP*I*JF$EJ ,EPG***E )EP6K*GG )EPEL6LH ,EPEHHKK )EPEE6LJ ,EP*GEG* ,EP6II*F$EK ,EP6KIFL ,EPFKL** ,EPEIFHI ,EP6H6LI ,EPH6F*6 )EPG6LIH )EP*E***$EL ,EPGGHGK )EPEJI6E )EPEJIEL )EPEG6KL ,EPEIHII ,EP*GJKH ,EP6HJKH$6E ,EP*6JJ6 ,EPEGIIK ,EPHIGL* ,EPHEHGH ,EPEKIJI ,EPEGHLL ,EPEEKKF$66 ,EP6JE*F ,EPFHHFL ,EPEIJHJ ,EPF6*6E )EPII6FI )EP*H6GL ,EPFJG*J$6* )EPE6JEL ,EPG*JEK )EP*JKH* ,EPF*KHG ,EP6IKHH ,EPHGFKK )EP6G6FFabcdeO *K 6ERS+ , $E6 66 , $E* *L , $EF , $EG FE , $EH 6F ^ 6* *J , $EI F6 F , $EJ $EJ $EL *E ` , $EK L $6E * , $EL *G I 6K$E* , $6E $EI , $66 *H $EH $66 $EK J , $6* *F *6 ** G $EG HK $E6*I[ QRS6, FF $6* 6G QRS*, $EF F* 6I] QRSF, 6H6L , QRSG, FG 6J 6 _ , QRSH, FH , QRSI, , QRSJ, , QRSK, , QRSL, , QRS6E, , QRS66, , QRS6*, Results
  28. 28. Method for Analysis• Principal Component Analysis (PCA) is done to behavioral data collected from the SFR task • all data are averaged • data showing over-variance are managed (over-variance suggests multiple attractors)• It was to see • if there are any determinants, i.e. Principal Components (PCs), that account for the patterning in positive evidence. • to represent, in terms of locations in a reduced, multi- dimensional space, how relevant situations are interrelated to each other, giving you some measures of the distances among them.
  29. 29. PCA Plot: President –real and nonce assassin man –words for Victim, sharp pain man - cancer old man - children - panic(condition C1), wasps woman – young man – strong jealousy pervertmeasuring zebras - lionsstrengths of passengers-by – lunatic mansituation country – family – runaway truckevocation by armed countryHarm-causer- Text city - influenzadenoting nouns bank branch – masked man town – gust of wind stock market - debacle+ at the center is the“neural” point, where NW- area - hurricanega NW-ni osowareta (“NW{was attacked by; was hit by;suffered (from)} NW”) islocated. 29
  30. 30. PCA Plot: President – assassinreal and nonce man – sharp painwords for Victim, children - man - cancer old man - panic(condition C1), wasps woman – pervert young man – strong jealousymeasuring zebras - lionsstrengths of passengers-by – lunatic mansituation country – family – runaway truckevocation by armed countryHarm-causer- Text city - influenzadenoting nouns bank branch – masked man town – gust of wind stock market -Arrows indicate debaclecorrespondences between area -nonce word ratings and real hurricaneword ratingsThickness of arrowsencodes strength 30
  31. 31. city - man - influenza cancer children - waspsPCA Plot: zebras - lionsreal and nonce old man -words for Harm- panic country – armed country man – sharp paincauser (condition area – PresidentC2), measuring hurricane woman – –assassin pervertstrengths of town – gust of wind passengers-by – bank branch – family – lunatic man masked man runaway trucksituation young man – strong jealousy old lady – purse snatcherevocation byVictim-denoting stock market –nouns debacleCompared to C1,attraction is weak, andlong 31
  32. 32. city - man - influenza cancer children - waspsPCA Plot: zebras - lionsreal and nonce old man -words for Harm- panic country – –armed armed country man – sharp country paincauser (condition area – President President –assassinC2), measuring hurricane woman – woman – pervert –assassin passengers-by – pervertstrengths of town – gust of wind passengers-by – lunatic man bank branch – family – lunatic man masked man runaway trucksituation young man – strong jealousy old lady – purse snatcherevocation byVictim-denoting stock market stock market –nouns –debacle debacleCompared to C1,attraction is weak, andlong 32
  33. 33. Results• Certain nonce words w* are feature-rated very much like real words, showing that their occurring contexts C(w*) = W(s) – w* evoke situations strong enough.• This confirms our prediction, and A-to-S Hypothesis is not falsified.• But • Clearly, different nouns have different strengths of situation-evocation • Some nouns (e.g., ansatu-sha ‘assassin’) showed stronger evocation effect; others (e.g., fukumen-no otoko ‘a masked man’) don’t show so much effect 33
  34. 34. Conclusion
  35. 35. A-to-S Hypothesis Confirmed;Yet ...• Overall, stronger A-to-S effect was found for Harm- causer nouns than for Victim nouns • because nouns for Harm-causer evoke situations stronger than nouns for Victim when used with osou?• Different nouns have different strengths of situation evocation • This supports the hypothetical distinction between role names from object names (Kuroda and Isahara 2005) • cf. Gentner (2005)’s relational nouns and object/entity nouns distinction 35
  36. 36. Why Distinguish Role Names from Object Names?• Nouns like victim, robbery, prey, predator, disaster, are role-denoting nouns distinguished from object- denoting nouns like (a) man, (a) typhoon• Instantiation relation (i.e., IS-A relation) is definable between object-denoting and role-denoting nouns: • [typhoon] IS-A [disaster], [flood] IS-A [disaster], etc. • [three people (wounded in the accident)] IS-A [victim]• This is a piece of information missing in most thesauri.• Not surprisingly, role names play more important a role in metaphorical mappings than object names (Nakamoto, Kuroda, and Kusumi, under review) 36
  37. 37. General Remarks• Certain nouns, if not all, evoke situations or frames. • probably independently from verbs and prepositions • or collaboratively with verbs and prepositions?• Situation evocation by nouns (direct evocation by role-denoting nouns; indirect evocation by object- denoting nouns), needs to be taken care of if we want to deal with noncompositional phenomena, including metaphor, successfully. • And a detailed specification and the successful description of it will supplement co-compositional processes given qualia structures (Pustejovsky 1995). 37
  38. 38. Acknowledgments• Toshiyuki Kanamaru, Kyoto University, NICT• Jae-ho Lee, NICT• Masao Utiyama, NICT 38
  39. 39. Thanks for Your Attention and Your Tolerance for My Not-So- Good English
  40. 40. After-thanks Slides
  41. 41. PCA Factor Loadings Feature ID Feature translated in English PC1 PC2 PC3 if anyVictim05 Y was aware of the possibility of victimization by X. -0.648 0.333 0.046Victim02 Y had a good chance to prepare for Xs activity. -0.641 0.199 0.183Victim09 The degree of Ys affectedness is greater than the individual scale. -0.555 -0.398 -0.166Victim08 Y could avoid Xs affect on it. -0.537 0.397 0.116Victim11 Y has been targeted by X long before. -0.535 0.303 -0.196Victim01 Y is a living thing. 0.484 0.725 -0.153Victim04 Y is human. 0.386 0.676 -0.088Victim03 Y had some reason to be victimized by X. -0.366 0.551 0.278Victim12 Y itself might have invited Xs activity on it. -0.442 0.550 0.207Victim07 Y is the name for a place. -0.549 -0.664 -0.018Victim06 Xs activity on X may cause X to die. -0.134 0.313 -0.746Victim10 Y suffered a harm by Xs affect on it. -0.286 -0.029 -0.713 Variance explained 2.826 2.686 1.333 % of variance 0.236 0.224 0.111 cummulative % 0.236 0.459 0.570 Interpretations PC1: Unpreditability of harm from X PC2: Number of sufferers (Scale)Harm-causer03 X is visible. 0.829 0.089Harm-causer02 X chose Y for its target. 0.828 0.029Harm-causer08 X did so to satisfy its desire or needs. 0.813 -0.055Harm-causer01 X is a living thing. 0.805 0.229Harm-causer05 X is human. 0.779 -0.163Harm-causer09 X planned to take off something from Y. 0.721 -0.037Harm-causer07 X is a natural phenomenon. -0.693 0.313Harm-causer11 Xs activity can kill Y. 0.238 0.669Harm-causer12 X is a collection of living things. 0.286 0.659Harm-causer10 X is the name for a sickness. -0.388 0.485Harm-causer04 X couldnt help doing it to Y. -0.134 0.196 Variance explained 4.596 1.345 % of variance 0.418 0.122 cummulative % 0.418 0.540 Interpretations PC1: Intendedness of X PC2: Seriousness of Ys suffering 41

×