Concept Composition

FCSS — Methodology & Concept Composition

Methodology

In the paper of Sebastian Löbner [Loeb2011], page 1, a systematic determination of concept types has been elaborated. Concepts are used as the representation of an abstract or concrete thing of the real world in the mind of a human. In linguistics and ontologies, the naming and use of concepts often becomes cumbersome due to the different meanings words can have, e.g. class, bank, ball, mouse etc. This may result in polysemy as discussed by Habibi et al [HaHa2021]. The approach presented here combines different methods to overcome these shortfalls. We show that concepts as multi word units can be composed from sub concepts, using Intra Concept Relationships (ICR) from the set ICR = {◊is, ◊Def, ◊Gen, ◊Dif, ◊Subject, ◊Relation, ◊Object}. Textual Definitions of a Word (TDW) or a concept in natural language – also called glosses - are replaced by precise formal Concept Binary Tree (CBT) definitions. Semantic Relations (SRs) between concepts are modeled with ICRs and with Extra-Concept Relationships (ECR): With Δhypernym = Δinv_hyponym = (inv, hyponym), Δhyponym = Δinv_hypernym = (inv, hypernym), Δholonym = Δinv_meronym = (inv, meronym), Δmeronym = Δinv_holonym = (inverse, holonym), Δtroponym = Δhyponym_verb = (hyponym, verb), we have ECR = {equal, inverse, antonym, hypernym, hyponym, holonym, synonym, troponym, homonym}. Concept Relationships (CR) are defined as the union of intra and extra SRs: CR = ICR ∪ ECR. Our concept construction patterns only use CBTs to compose Semantic Compounds (SC) as super concepts from sub concepts or Semantic Primes (SP).

Concept Composition

To tackle the problem of untranslatability of concepts identified in [GoWi2015a], our approach is based on the Minimal English Lexicon (MEL) and Longmans Defining Vocabular (LDV). In this section, methods and guidelines are introduced for the composition of the different word types nouns, verbs, adjectives, and propositions. Willem Levelt describes in [Leve1989], page 288 as important restriction that many of the word types can be combined, but some not, such as (noun, verb) or (verb, adjective). Many of the following patterns are derived from the work of Löbner  [Loeb2011].

  Compound Nouns:
Compound NouNs (CNN) are constructed with the Genus Differentiae Pattern (GDP) and define a hypernym relationship, for example >BLS-air_pressure = (Δ>BLS-pressure, Δ>BLS-air) = (pressure, air). Many Sortal NouNs (SNN) take naturally adjectives like Δtall = (adj, tallness). Individual NouNs (INN) include proper names and personal pronouns along with nouns for unique institutions: Pope, Paris, weather(Munic), temperature(Sun);  Relational NouNs (RNN) are binary predicate terms where the object is usually specified by a possessive construction and therefore call possessor. Examples of RNNs include most kinship terms and a broad variety of Deverbal NouNs (DNN). Here are some examples for kinship and for functional axiom / rule definitions:
  • ωgrandfather_maternal = (>BLS-grandfather_(maternal), »Def, (father, mother));
  • ωgrandfather_paternal = (>BLS-grandfather_(paternal), »Def, (father, father)).
This is extended by the hypernym definitions Δgrandfather_paternal = (is, grandfather); Δgrandfather_maternal = (is, grandfather) = Δis_grandfather. But please note, the definition Δgrandfather = Δfather_father has to be regarded as an anti-pattern, because it would conflict with Δgrandfather = Δfather_mother and be the source of errors in reasoning.
  Functional Nouns:
Functional Nouns (FNN) are unary function terms. Their meanings are functional concepts, involving the possessor as argument. Examples include role terms like Δfather(Rosi) = (father, Rosi), and Δpresident(France) = (president, France), as well as terms for unique parts such as Δcover(computer) = (cover, computer) and also terms for abstract aspects like Δname(Pope) = (name, Pope), dimensions like Δage(Picasso) = (age, Picasso), or Δmeaning(meaning) = (meaning, meaning) .
  Conceptual Nouns Phrases (CNP):
CNP = (determiner, noun): CNPs are characterized by modes of determination, e.g., simply indefinite: (a, table), unspecific indefinite: (some, people); free choice: (any, knife); negative indefinite: (no, employee), contrastive demonstrative: (this, car), (that, painting); plural: women = (plural, woman); numerical: Δcentury = Δhundred_years = (hundred, year); quantitative: (many, people), (several, products), (some, kilogram); quantificational: (every, hour), (each, person), (all, country), (both, side). CNPs go naturally with singular definite determination, e.g., definite articles indicate a conceptual type of reference like in (the, car), (the, plan).
  Prepositional Concepts (PPC):
 are composed from a preposition and a noun with PNP = (prep, noun) such as in Δunder_water = (under, water), Δbehind_Moon = (behind, Moon), Δthrough_air = (through, air), Δinside_house = (inside, house), and Δinto_nothing = (into, nothing). According to [Loeb2011], uniqueness is a semantic property of individual and functional nouns while relationality is a semantic property of relational and functional nouns as in Δtooth_of_a_dog = (tooth, (a, dog)).
  Possessive / Property Chains:
The example from[Loeb2011] „my father’s wife’s mother’s car“ can be represented as CBT car(mother(wife(my, father))). CBTs can be used for inferencing as follows: With Rosi ≡ Δmother_Mary = (mother, Mary) and Tom ≡ Δfather_Rosi = (father, Rosi) → Tom = (father, Δmother_Mary ) = Δfather_Δmother_Mary = father(mother(Mary)). With grandfather_(maternal) = (father, mother) ≡ father(mother), the property chain matches and leads to the inference Tom ≡ (grandfather(paternal), Mary) and with grandfather_(paternal) = (is, grandfather) follows Tom ≡ (instance, grandfather). This inferencing / entailment result could then be stored as an assertion.
  Activities:
The pattern for the construction of compound concepts from activities is ({verb, adjective, result, cause}, activity): Adjectives of activities are constructed with the Genus Differentiae Pattern (GDP) and usually do not have comparatives and superlatives: Δprint_(to) = Δverb_printing = (verb, printing), Δprinted = Δadj_printing = (adj, printing), Δprint = Δproduct_printing = (product, printing).

Latter corresponds to the definitions (>BLS_printing_product, »Gen, >BLS_product) and (>BLS_printing_product, »Dif, >BLS_printing).

  Opposites:
When defining concepts with the semantic relations inverse (inv), opposite_of (opp), and perpendicular (perp / orthogonal), as a general guideline we propose to take that concept as a base concept which is most large, high, on_top etc. In the case where an orientation or direction is involved, like in north, east, south and west, the suggestion is to go clockwise to choose the orthogonal/vertical concept like Δsouth = Δopp_north = (opp, north); Δeast = Δperp_south = (perp, south) etc.
  Summary of Binary Tree Compound Patterns:
Compound and Functional noun definitions take the form NNP = (noun, noun), Conceptual Noun Phrases CNP = (determiner, noun), Adjective and Adverbial Object AAC = ({adj, adverb}, concept), Prepositional Concepts PNP = (prep, noun), Adjectives of Quantities and Qualities AQQ = (adj, noun), Comparatives CMP = (more, adj), Superlatives SUP = (most, adj), Opposites OPP = (opp, concept), and a recursively defined Possessive / Property Chains PPP = (noun, PPP). A PPP can be chain of n nouns, for example, Name(Wife(Grandfather_(maternal)) ≡ (Name, (Wife, Grandfather_(maternal))) and in general for n=3 we have x_of_y_of_z = x_y_z = (x, (y, z)).

Extension: deriver.app

Combined mirror of FCSS Methodology and FCSS Concept Composition. Back to Controlled vocabulary of concepts; Deriver documentation.

Sources: FCSS Methodology, FCSS Concept Composition.