TITULO III DE LAS DEDUCCIONES
GANANCIAS DE FUENTE EXTRANJERA OBTENIDAS POR RESIDENTES EN EL PAÍS
The CoreLex resource (Buitelaar, 1998) defines a set of 39 basic types. These basic types are assigned to noun synsets in WordNet and are meant to indicate a property of all subordinate synsets. For example, the basic types animal or food are assigned to different WordNet vertices and thus define the type of these vertices and their hyponyms. All
synsets that are a child element of animal are animals and all child elements food are foods.
As mentioned before, nouns are hierarchically ordered and the topmost (empty) synsets are called unique beginners. The basic types are an extension to the existing hyper- nymy/hyponymy tree of the WordNet noun subnet and aim to be more exact and espe- cially useful for the identification of regular polysemy.
Many of the approaches presented here are based on, or are related to, CoreLex and thus restricted to nouns. The general idea is to provide a taxonomy of basic types that can be used to identify regular polysemy based on the definition by Apresjan (1974) seen before. Taking for example the grinding rule, it is expected to find regular patterns of co-occurrence of homograph words both in the set of vertices subsumed under animal and food.
Based on the CoreLex resource, Boleda et al. (2012) and Utt and Pad´o (2011) present
an approach to identify regular polysemy (i.e., common patterns used by speakers that form polysemous words such as in grinding). The authors identify common pairs of basic types, like the one shown above, that indicate a productive (i.e., regular) polysemy pattern
(cf. Utt and Pad´o, 2011). If given a word like lamb with many word forms in WordNet,
the system looks up the sense’s basic types. If it finds one sense corresponding to the type animal and another of the type food, this pattern can also be found for chicken and others.
Utt and Pad´o (2011) state that “[i]n polysemy . . . sense variation is systematic, i.e., ap-
pears for whole sets of words” (Utt and Pad´o, 2011, p. 265) and go on to define polysemy,
with respect to the mentioned 39 basic types, as “high-frequency (systematic) basic ambi- guities . . . , and low-frequency (idiosyncratic) basic ambiguities as homonymous” (ibid.). The basic types a lemma is connected to are expected to show patterns of reoccurring combinations of basic types indicating regular polysemy. The more of these patterns a lemma is related to, the more polysemous it is.
Even though in the context of the paper this definition is an understandable reduction, it is, in my opinion, problematic in many aspects. Polysemy is not only regular polysemy which is what is in fact defined here. In many cases more subtle or deep relations between two senses play a role that cannot be explained by these regular patterns.
Taking the lamb example, lamb can also mean “a sweet innocent mild-mannered person (especially a child)” (WordNet), which refers not to the fact that lamb is an animal or
polysemy as a property of words, not of word forms. If among the word forms they find at least two that fit one of the defined patterns, this word is treated as polysemous. This is in my opinion an oversimplification and should be avoided. There are cases in which some word forms belonging to one lemma are indeed related while others belonging to the same lemma are obviously not (e.g., bank may refer to a financial institution and the building that institution resides in (polysemy) as well as to a slope of land (homonymous)). To
avoid this pitfall, or to at least reduce the negative effect, Utt and Pad´o (2011) calculate
a ratio whose values lie between 0 and 1.
Afterwards they define a cutoff value of the ratio >0 to distinguish between polysemous
words and homonymous words. Utt and Pad´o (2011, p. 268) state that they “consider
polysemy and homonymy the two points on a gradient, where the words in the middle show elements of both”. Although this is a nice thought, it of course leads to nothing since they still define a border (i.e., cutoff value) between both. Also the necessity to define such a ratio in the first place is only necessary because polysemy is defined on the word level and not on the sense level. The question remains how to treat the different senses and related word forms of bank : In WordNet, most of the senses of bank are connected to the financial institution, only four to mound or river bank. Is bank thus polysemous since most senses are related, or is it homonymous since there are two quite different clusters of meaning? Many people would surely agree that bank is homonymous. Regarding the problem of identifying word forms that are related in WordNet, and WordNet is based on senses, not on words or lemmas, neither of the two possibilities solves the problem. The only possibility is to adapt the problem and the definition of homonymy and polysemy to WordNet and look at the relations on the level of word forms. Also, when thinking of the possible use-case in word WSD, working with sense instead of words is necessary since the task is to distinguish between the different senses based on the usage of a word.
Peters and Peters (2000) examine how the-top level vertices (nouns.TOPS ) of the WordNet hierarchies could be used to identify regular polysemy. The top-level vertices each enfold a tree of hypernym/hyponym relations beneath it. The idea is similar to that of CoreLex but uses the structure of the WordNet data set and uses the different files. Each file corresponds to one top-level vertex used to store parts of the noun tree.
Every noun in WordNet is a child of at least one of these unique beginners, i.e., categories that structure the whole noun subgraph (e.g., artifact: a man-made object or group: any number of entities (members) considered as a unit ). Peters and Peters (2000) identify cases where polysemous lemmas are members of both artifact or group.
For example, the name of an institution can refer to an organization (group) and the corresponding building (artifact ). This also shows the metonymic character of regular polysemy since it can be argued that the building – belonging to, owned, or rented by the institution – is part of that institution.
To identify more cases, they also added other hypernyms within a given geodesic path (≤ 4) to find pairs of other (super) concepts that show a similar regularity. So taking a
word form w1 and word form w2 that share the same lemma, not only their corresponding
unique beginner is taken into account, as well as their hypernyms up to the geodesic path of 4.
The polysemy concept used by Peters and Peters (2000) is similar to the one used here as it only looks at pair-wise combinations of word forms and at words, i.e., sets of word forms that correspond in appearance but necessarily in meaning, to identify polysemy. It is the internal WordNet structure, the non-existence of the word concept, and the organization in synsets formed by word forms that makes this treatment of polysemy a property of word forms, and not of words, plausible in this context.
The goal of the Peters and Peters is to extend WordNet by looking for missing pairs (i.e., cases where one sense is missing). One pair is that of container/quantity (e.g., cup is the container as well as the amount of substance in it). The same pattern of regular polysemy could be applied to every container (e.g., an amphora) to build the sense of the quantity, as in He drank the whole amphora of wine. According to this goal, their evaluation only looks at those cases that were identified to see if they are found in WordNet and thereby right (true positives), or wrong (true negatives) (i.e., those pairs that did not shown the pattern) without taking into account those cases that were not covered by their
approach (false positives) (i.e., patterns that might not be identified in the first place).86
Another interesting, diverging, method is introduced by Veale (2004). Instead of
looking at structural similarities, he proposes a system that compares the glossaries of synsets containing potentially polysemous word forms to identify those that are in fact related. Since this method is not applicable in the context of this thesis (i.e., using the network structure to identify missing relations) because it ignores the structure of WordNet altogether, I will not elaborate on the details of how this similarity is computed. My approach differs from the ones cited above. First, after looking at the different approaches and at the structure of WordNet, I consider polysemy a relation on the level
86Another approach that is based on the systems mentioned so far is presented by Barque and Chau-
of word forms (i.e., the elements that make up a word sense) and not on the level of words, a concept that does not exist in WordNet in this form. Some word forms can be seen as polysemous: their sense is related and they share a lemma, while other forms only share a lemma but the senses are not related. There is no binary opposition on the word level between polysemous and homonymous words but only between related (or polysemous) and unrelated (or homonymous) pairs of word forms. A word can therefore exhibit polysemous senses (cf. to bank (money) and bank as financial institution) and homonymous senses (cf. the before mentioned senses of bank and one meaning shore or sandbank) at the same time. Some senses of bank form a set of polysemous semantically related meanings while at the same time they share a common form (homograph) with others only by chance. Secondly, I do not restrict my approach to only nouns. Restricting the research to only nouns seems to be a result of two factors. One is the hierarchical structure of the noun graph that supports disambiguation. CoreLex is based on nouns and often used. Furthermore, the hierarchical structure makes computing the similarity easier than in other subnets that do not offer hierarchies. The other reason is the fact that many researchers in the field have information retrieval tasks in mind and nouns are by far the most searched for word class. But the distinction between polysemous and homonymous might also play a role in word sense disambiguation and machine translation tasks where there is no restriction to only nouns.
Also, the approach presented here does not ignore irregular cases of polysemy. The goal is to identify both those cases that are formed by regular rules, as well as those formed using some other kind of similarity between object, be this relation metaphorical or metonymic.