Servicization of language resources in a Web-based environment has opened up the potential for dynamically combined virtual lexical resources. Evolving lexical linked data could be realized, provided being recovered/discovered links among lexical resources are properly organized and maintained. This position paper examines a scenario, in which lexical semantic resources are cross-linguistically enriched, and sketches how this scenario could come about while discussing necessary ingredients. The discussions naturally include how the existing lexicon modeling framework could be applied and should be extended.
Servicization of language resources provides the potential of a dynamic lexical
resource [
This position paper argues that by opportunistically associating different lexical resources across a language barrier, relevant portion of the lexical resources can be gradually enriched and could be made public by standing on the Linked Data mechanism. This paper also argues more relationships could be acquired, when there exists a lexical semantic disparity. The presented work concentrates on WordNet-type semantic lexicons. Their fundamental information structures are represented by the following lexical class objects. { A Lexical Entry comprises of Forms and Senses. { A Form can be a Lemma or a Phrase; the latter comprises of more than one
Lemmas. { A Sense denotes a Synset. { A Synset is denoted by one or more Senses.
{ Synsets are linked by one of the predefined Conceptual Relations. 3
Below we introduce a motivating example, where an English query term gadget is issued to search for a set of corresponding Japanese translations, each hopefully grounded in a Japanese conceptual system. Suppose we get two translations, under the same sense division, for gadget by using an appropriate translation resource: t1 :”” (gajetto ), which is the transliteration of gadget, and t2 :”” (yuuyounakiki ), which actually is a two-word phrase. 3.1
“yuo”na :2t “yuonaki” c2 a2 a1 into [ (yuuyouna)/Adj, (kiki)/Noun], a Phrase node is introduced to associate this two-word phrase with its constituents by the c1 and c2 links.
These successive operations are invoked directly while handling the query; we thus call them direct linking of lexical objects. Note that the ad-hoc Synset node is yet to ground in the Japanese conceptual system at this time. 3.2
While the structure around t1 has been settled in the current configuration, that of around the ad-hoc Synset node for t2 can be further enriched, again by seeking cross-lingual correspondences. Figure 2 summarizes the outcomes. “yuo”na
Two cross-lingual synset-to-synset links (d1 and d2 ) are first introduced by associating a sense of ”” (kiki ) with a sense of device and a sense of ”” (yuuyouna) with a sense of useful respectively. By establishing d1, the semantic head of the ad-hoc synset for t2 is then identified and represented by the link e1. The same story holds for the semantic modifier of t2, and the link e2 is introduced to represent this semantic relationship. These operations also enable the introduction of the link f, which, in a sense, shows ”” (gajetto ) can be rephrased as ”” ( yuuyounakiki ).
The evolving story so far signifies us the possibility of lexical knowledge enrichment that takes advantage of the opportunity to interrelate lexical objects across a language barrier. Let us remind that a semantic gap brought about by differences in the lexicalization would provide us a further opportunity to enrich relevant range of the existing lexical structures.
We could acquire more correspondences as illustrated in Figure 3 by further pursuing this strategy. In the figure, another ad-hoc Synset node in the English lexical space, and two semantic links (g1 and g2 ) to label the semantic head/modifier of the ad-hoc synset are introduced. Besides, the ad-hoc Synset node is linked to that of gadget by the link h; this is in parallel with the link f in the Japanese lexical space. Notice again that almost instant introduction of these links is originated from the cross-lingual synset-to-synset matching that is invoked for establishing the correspondences represented by d1 and d2. “yuo”na :1t oe”tjag“ c1 J: ivtcejda tpenoc 1e a2 a1 d2
We would call these secondary operations initiated after the direct linking as indirect linking. The lexical objects introduced in this motivating example are examined in more detail in the next section to sort the necessary elements to realize the scenario.
The basic lexicon model described in section 2 has to be extended in some ways.
First, in the motivating example, two ad-hoc Synset nodes were introduced to accommodate the two-word translation phrase t2, and the corresponding virtual phrase (could be verbalized as useful device) in English. These nodes, in their nature, may be ad-hoc and represent a kind of complex concept that may lexicalize to a phrase rather than a single word in one language. Therefore an instance of the ad-hoc Synset class should have an attribute to indicate the instance is typed complex, and could have Morpho-syntactic Head/Modifier links (like c1,c2 ) as well as Semantic Head/Modifier links (like e1,e2,g1,g2 ).
Second, some of the introduced links should be typed differently from the
existing lexicon model. Table 1 classifies the links introduced in the motivating
example. The link type #1 is of intrinsic important in the presented
framework. As the correspondence between synsets in different languages, in a sense,
is rarely equivalent [
The link type #3 represents morpho-syntactic head/modifier relationships, whereas link type #4 represents semantic head/modifier relationships. As far as semantic compositionality holds, these two link types exhibit a kind of parallel structure as illustrated in the example: the semantic links (e1 and e2 ; typed #4) were eventually introduced, corresponding to the already existing morphosyntactic links (c1 and c2 ; typed #3).
On the other hand, in cases where the semantic compositionality does not hold, we should demur the introduction of these semantic links, even each of the Japanese synsets could find their mates in the English lexical space. In such a case, we have to devise an independent method to check the semantic compositionality, or we should seek more semantic constraints to apply, probably from the English lexical space; but this issue largely remains as a future issue.
As for the actual modeling and representation of lexical resources, we can
rest with the existing frameworks, including the ISO standard lexical markup
framework (LMF) [
One of the most important elements is obviously a computational process for finding a synset mate in another language. We are now studying a method to calculate semantic similarity between synsets across a language, by simply employing bilingual translation resources and probability distributions acquired from a sense-tagged corpus in the target language.
We can also apply and/or combine previously proposed methods. For
example, the method reported highly accurate [
The following issues have to be considered in implementing an effective operating environment. First, we need to have a global mechanism to control the indirect linking operations. As shown in the example, indirect links can be introduced upon establishment of a direct link. However who/what should decide to initiate the indirect linking process is unclear. Moreover, to what extent the indirect linking should be propagated remains uncertain. Second, we are in need of having a proper vocabulary to annotate the lexical objects that participated in direct/indirect linking operations. For example, we would need to know when and how a particular link was established. We thus need to have a sort of ontology for describing linking events, which naturally includes references to the linguistic processes that were actually applied, as well as the human approvals.
This position paper presented a notion of evolving linked data, in which recovered/discovered relationships among lexical objects would be published as links. It also argued that the associated lexical resources could be enriched further, in particular cases where a sort of lexical semantic disparity exists.
Acknowledgments. The presented work was supported by KAKENHI (21520401) provided by MEXT, Japan, and the Strategic Information and Communications R&D Promotion Programme (SCOPE) of the Ministry of Internal Affairs and Communications of Japan.