| Antonymy and Conceptual Vectors |
| Outline |
| The main idea | |||
| Background on conceptual vectors | |||
| How we use CVs | |||
| & why we need to distinguish CVs of antonyms | |||
| Brief study of antonymies | |||
| Representation of antonymies | |||
| Measure for antonymousness | |||
| The main idea |
| Work on meaning representation in NLP,
using conceptual vectors (CV) |
|||
| applications = WSD & thematic indexing | |||
| but V(existence) = V(non-existence) ! | |||
| basic concepts activated the same | |||
| Idea: | |||
| use lexical functions to improve the adequacy | |||
| For this, transport the lexical functions in the vector space | |||
| Background on conceptual vectors |
| Lexical Item = ideas = combination of
concepts = Vector V |
||
| Ideas space = vector space (generator space) | ||
| Concept = idea = vector Vc | ||
| Vc taken from a thesaurus hierarchy (Larousse) | ||
| translation of Rogets thesaurus, 873 leaf nodes | ||
| the word peace has non zero values for concept PEACE and other concepts | ||
| Our conceptual vectors Thesaurus |
| H : thesaurus hierarchy K concepts | ||
| Thesaurus Larousse = 873 concepts | ||
| V(Ci) : <a1, , ai, , a873> | ||
| aj = 1/ (2 ** Dum(H, i, j)) | ||
| Conceptual vectors Concept c4: PEACE |
| Conceptual vectors Term peace |
| Diapositive 8 |
| Angular or thematic distance |
| Da(x,y) = angle(x,y) = acos(sim(x,y)) | ||
| = acos(x.y /|x ||y |) | ||
| 0 D(x,y) p (positive components) | ||
| If 0 then x and y are colinear : same idea. | ||
| If p/2 : nothing in common. | ||
| Thematic Distance (examples) |
| Da(anteater , anteater ) = 0 (0) | |||
| Da(anteater , animal ) = 0,45 (26) | |||
| Da(anteater , train ) = 1,18 (68) | |||
| Da(anteater , mammal ) = 0,36 (21) | |||
| Da(anteater , quadruped ) = 0,42 (24) | |||
| Da(anteater , ant ) = 0,26 (15) | |||
| thematic distance ontological distance | |||
| Vector Proximity |
| Function V gives the vectors closest to a lexical item. | |
| V (life) = life, alive, birth | |
| V (death) = death, to die, to kill |
| How we build & use
conceptual vectors |
| Conceptual vectors give thematic representations | ||
| of word senses | ||
| of words (averaging CVs of word senses) | ||
| of the content ( ideas ) of any textual segment | ||
| New CVs for word senses are permanently learned from NL definitions | ||
| coming from electronic dictionaries | ||
| CVs of word senses are permanently recomputed | ||
| for French, 3 years, 100000 words, 300000 CVs | ||
| Continuous building of the conceptual vectors database |
| We should distinguish CVs of different but related words |
| Non-existent : who or which does not exist | |||
| cold : #ant# warm, hot | |||
| Without a specific treatment, we get | |||
| V(non-existence) = V(existence) | |||
| V(cold) = V(hot) | |||
| We want to obtain | |||
| V(non-existence) V(existence) | |||
| V(cold) V(hot) | |||
| in order to improve applications and resources |
| Applications: more precision | ||
| Thematic analysis of texts | ||
| Thematic analysis of definitions | ||
| Resources: coherence & adequacy | ||
| General coherence of the CV data base | ||
| Conceptual Vector quality (adequacy) | ||
| Lexical functions may help! |
| Lexical function (Meltchuk): | |||
| WS {WS1WSn} | |||
| synonymy (#Syn#), antonymy (#Anti#), intensification (#Magn#) | |||
| Examples : | |||
| #Syn# (car) = {automobile} | |||
| #Anti# (respect) = {disrespect; disdain} | |||
| #Sing# (fleet) = {boat, ship; embarcation} | |||
| Method: transport the LFs as functions on the CV space |
| e.g. for antonymy, | |||
| to get V(non-existence) V(existence) | |||
| find vector function Anti such that: | |||
| V(non-existence) | |||
| = V(#Anti#(existence)) = Anti (V(existence)) | |||
| similarly for other lexical functions | |||
| we simply began by studying antinomy | |||
| Brief study of antonymy |
| Definition : | ||
| Two lexical items are in antonymy relation if there is a symmetry between their semantic components relatively to an axis | ||
| Antonymy relations depend on the type of medium that supports symmetry | ||
| There are several types of antonymy | ||
| On the axis, there are fixed points: | ||
| Anti (V(car)) = V(car) because #Anti# (car) = | ||
| 1- Complementary antonymy |
| Values are boolean & symmetric (01) | |
| Examples : | |
| event/non-event dead/alive | |
| existence/non-existence | |
| He is present He is not absent | |
| He is absent He is not present |
| 2- Scalar antonymy |
| Values are scalar | ||
| Symmetry is relative to a reference value | ||
| Examples : cold/hot, small/tall | ||
| This man is small Þ This man is not tall | ||
| This man is tall Þ This man is not small | ||
| This man is neither tall nor small | ||
| reference value = of medium height | ||
| 3- Dual Antonymy (1) |
| Conversive duals | |
| same semantics but inversion of roles | |
| Examples : sell/buy, husband/wife, father/son | |
| Jack is Johns son John is Jacks father | |
| Jack sells a car to John John buys a car from Jack |
| 3- Dual Antonymy (2) |
| Contrastive duals | ||
| contrastive expressions accepted by usage | ||
| Cultural : sun/moon, yin/yang | ||
| Associative : question/answer | ||
| Spatio-temporal : birth/death, start/finish | ||
| Coherence and adequacy of the base |
| Learning bootstrap based on a kernel composed of pre-computed vectors considered as adequate | |
| Learning must be coherent = preserve adequacy | |
| Adequacy = judgement that activations of concepts (coordinates) make sense for the meaning corresponding to a definition | |
| For coherence improvement, we use semantic relations between terms | |
| Antonymy function |
| Based on the antonym vectors of concepts : one list for each kind of antonymy | |
| Antic (EXISTENCE) = V (NON-EXISTENCE) | |
| Antis (HOT) = V (COLD) | |
| Antic (GAME) = V (GAME) | |
| Anti (X,C) builds the vector opposite of vector X in context C | |
| Construction of the antonym vector of X in context C |
| The method is to focus on the salient notions in V(X) and V(C) | |
| If the notions can be opposed, then the antonym should have the inverse ideas in the same proportions | |
| The following formula was obtained after several experiments |
| Construction of the antonym vector (2) |
| AntiR (V(X), V(C)) = Pi *AntiC (Ci, V(C)) | ||
| Pi = V * max (V(X), V(Ci)) | ||
| Not symmetrical | ||
| Stress more on vector X than on context C | ||
| Consider an important idea of the vector to oppose even if it is not in the referent | ||
| Results |
| Antonymy evaluation measure |
| Assess how much two lexical items are antonymous | |
| Manti(A,B) = DA(AB, Anti(A,C) Anti(B,C)) | |
| Examples |
| Manti (EXISTENCE, NON-EXISTENCE) = 0,03 | |
| Manti (existence, non-existence) = 0,44 | |
| Manti (EXISTENCE, CAR) = 1,45 | |
| Manti (existence, car) = 1,06 | |
| Manti (CAR, CAR) = 0,006 | |
| Manti (car, car) = 0,407 |
| Conclusion and perspectives |
| Progress so far : | |||
| Antonymy definition based on a notion of symmetry | |||
| Implemented formula to compute an antonym vector | |||
| Implemented measure to assess the level of antonymy between two items | |||
| Perspectives : | |||
| Use of the symbolic opposition found in dictionaries | |||
| Search the opposite meaning of a word | |||
| Study of the other semantic relations | |||
| (hyperonymy/hyponymy, meronymy/holonymy) | |||