In this section, we consider a more global level of lexical organization. Lexical semantics relations play an essential role in lexical semantics and intervene at many levels in natural language comprehension and production. They are also a central element in the organization of lexical semantics knowledge bases. Most of the material presented here is borrowed from [Cru86].
Two words W1 and W2 denoting respectively sets of entities E1 and E2, are in one of the following four relations:
There are basically three major types of hierarchical relations: taxonomies, meronomies and proportional series.
The taxonomy relation is the well-known isa relation which associates an entity of a certain type to another entity (called the hyperonym) of a more general type. Taxonomy intoduces a type/subtype relation which can be characterized by one of the following linguistic tests:
X is a subtype of Y if the following expressions are correct:
X is a kind of Y or X is a type of Y for nouns,
X-ing is a way of Y-ing for verbs.
Taxonomies usually have up to 7 levels that correspond to different levels of genericity (as in natural taxonomies). However, taxonomies of technical terms may be much deeper. It is also important to note that in some cases, certain nodes do not have any corresponding word in a given language; whereas they have one in another language. A taxonomy may thus have holes. Methods to define taxonomies are given in (3.4.3). The main property of a taxonomy is transitivity of properties from the type to the subtype. This property can also be viewed as a well-formedness criterion for taxonomies.
Most levels of a certain degree of genericity have a large number of subtypes, each of them having different possible realizations as words. The notion of subtype is however difficult to qualify in an homogeneous way. There is indeed a problem of prototypicality which is raised: some subtypes are more prototypical than others of their hyperonym (the type above them). Let us recall the famous example of the blackbird which is more prototypical of a bird than a hen which is itself more prototypical of that same class than a penguin.
Meronimies describe the part-whole relation. It is a fairly complex relation which attempts to take into account the degree of differentiation of the parts with respect to the whole and also the role that these parts play with respect to their whole. For example, elements such as spatial cohesion and spatial differentiation, functional differentiation and nature of the links between the parts are crucial elements for determining meronomies. In fact, depending on the quality of these elements, we may have different kinds of meronomies, with different types of properties.
Meronimies can be characterized perhaps in a slightly too restrictive way, by the following linguistic tests. A is a part of B if one of these sentences is correct:
B has A (or B has a A),
A is part of B.
The meronomy relation has itself some properties (or attributes) which must be taken into account in any realistic model:
Similarly to taxonomies, the meronomy relation cannot really be conceived between two elements, but should be concieved with respect to the set of all the parts forming the whole. This also permits to introduce a kind of point of view in a meronomic description. Meronomies do not, in general, allow transitivity at logical and linguistic levels. However, some authors tend to allow transitivity at linguistic level between elements which are linked by the same subtype of meronomic relation described above.
Non-branching hierarchies allow for the ordering of elements that correspond to different levels of organization or of dimensionality. The structure does not correspond to a type/subtype organization, but could have in somes cases some similarity with a meronomic relation. Non-branching hierarchies are often related to a spatial, a temporal or an abstract notion of dimensionality.
We can distinguish three kinds of non-branching hierarchies:
In some cases, non-branching hierarchies may reflect a more linguistic than common-world knowledge.
Among non-hierarchical relations we mainly distinguish synonymies and the different forms of opposition. These relations, as we shall see, are either binary or ternary. The ternary character reflects the context-dependence of some of these relations.
Two words are synonyms if they have a significant similar semantic content. Synonyms have a significant semantic overlap, but the degree of synonymy is not necessarily related to that overlap. There are very few absolute synonyms, if any, in a language, but words may be synonyms in given contexts. We then view the synonymy relation as a ternary relation : W1 and W2 are synonyms in the context C. Synonyms often do not depend on the degree of precision of the semantic descriptions, but their degree of synonymy may however change at different levels of granularity.
Antonyms and Opposites
Antonyms and opposites cover a very large variety of phenomena, more or less clearly defined. A basic definition could be that W1 and W2 are antonyms or opposites if they have most semantic characteristics in common but if they also differ in a significant way on at least one essential semantic dimension. As with synonyms, antonyms and opposites are highly contextual and thus introduce a kind of ternary relation. There are also various degrees of opposition: some pairs of word-senses are more prototypically opposites than others. Antonyms refer to gradable properties and opposites to non-gradable ones.
For example, with respect to the context `to start', to keep on and to stop are opposites. Similarly, good and bad are generally admitted as antonyms, and are more prototypical than the opposition between father and mother.
Antonyms do not necessarily partition the conceptual space into two mutually exclusive compartments which cover the whole conceptual domain. Some overlapp or space in between is possible, as in good and bad, since it is indeed possible to say that something is neither good nor bad, or, possibly, to say that something is both good and bad. A special class of antonyms are complementaries which divide the whole conceptual space into two non-overlapping compartments. In [Cru86] several classes of complementaries are defined, such as the class of interactives, which represent a relation of the type stimulus-response, as in: grant - refuse with respect to the current context.
Another interesting class among opposites are directional opposites. They represent either basic, topological, or conceptual (metaphorical) directional oppositions. In this class, which is conceptually relatively simple, fall examples such as: start-finish, top-bottom, descend-ascend.
The role of opposites in a lexical semantics knowledge base is somewhat difficult to define. Similarly to synonyms, opposites and antonyms may certainly play the role of integrity constraints. Their use in natural language generation, for example to avoid the use of too many negations, is somewhat hard to make explicit, because of numerous pragmatic factors that may intervene, such as the polarity of an element in a pair of opposites or antonyms. We can say, for example:
how expensive is this book ?
but probably not:
how cheap is this book ?
Finally, the linguistic tests or the analysis methods for defining exactly if two elements are opposites or antonyms and to what degree remain to be defined precisely.
Lexical semantic relations are of much use in structuring lexical data, particularly hierarchically. They have been extensively used and evaluated in WordNet 3.4.2 and EuroWordNet 3.4.3, they have also been used more experimentally in projects such as Acquilex 3.10.3 or Delis 3.10.5.
Lexical semantic relations are not used as such in applications, although one may use them in natural language generation 4.5 in lexicalization. They are mainly used to structure the lexicon 5.