Compound labels which have an entry in linguistic resources such as lexical databases, dictionaries, etc. (for example “jet lag”, “travel agent” and “bed and breakfast”) are treated as single words in our approach. Others like “railroad transportation”, which have no entry in the previous resources, are translated using a compositional method (see figure 6.17)
This approach uses a hybrid composition of two translation components. The first component is the same as the first component used for translating simple labels. The second component relies on a similar approach to the example-based method identified as part of the corpus-based translation techniques (see section 5.5). There are two main differences. First, instead of extracting the bilingual translation templates from a simple monolingual corpus or from parallel corpora, we derived these templates from different ontologies. The second difference relates to the used method to discover the translations. We do not extract the translations from a corpus, but from different linguistic resources.
Figure 6.17: LabelTranslator strategy to localize concept, attributes and relations represented by compound labels.
In a nutshell, this method splits the label into tokens (“railroad” and “transportation” in the example); the individual components are translated
and then combined into a compound label in the target language. Care is taken to combine the components respecting the word order of the target language. A set of lexical templates derived from different ontologies are used to control the order of translation. The main steps of the algorithm are:
1. The compound label is normalized, e.g., rewritten in lowercase, hy- phens are removed, it is split into tokens (see segmentation task in the figure), etc.
2. A set of possible translations is obtained for each token of the com- pound label using the different translation paradigms (first component in our translation strategy). The method uses all possible combina- tions of translation obtained for each token.
3. Since translations between languages do not keep the same word order, the algorithm creates candidate translations in the target language using lexical templates20. Each lexical template contains at least a pair
of patterns, namely ‘source’ and ‘target’ patterns. A source pattern is a template to be compared with the tagged compound label21, described in the source language, while the target pattern is used to generate the label in the target language. If no applicable template is found, the compound label is directly translated by the translation service. 4. All the candidate labels that fulfill the target pattern are returned as
candidate translations of the compound label.
In the following we describe the process to learn the lexical templates which are used to control the order of translation of compound labels.
Learning Lexical Templates from Ontological Labels
We believe that lexical templates used to translate compound labels are a necessary component to produce high quality translations because 1) it guarantees grammatical output and, 2) it makes sure that the structural source language meaning is preserved. In our approach, we used a semi- automatic process to obtain the lexical templates. As we explained before, each lexical template is composed of source and target patterns. The ontol- ogy labels used to learn the source patterns were extracted from different domain ontologies expressed in English, German, or Spanish. Each label was tokenized and tagged using the language independent part-of-speech tagger proposed in [TreeTagger, 1997]. The labels used to learn the target
20
The notion of lexical template proposed in this paper refers to text correlations found between a pair of languages.
21
We use TreeTagger [TreeTagger, 1997] in order to annotate the compound labels with part-of-speech and lemma information.
patterns were extracted either from the multilingual information associated with each ontological term or by means of a manual translation process. The same process used to annotate part of speech (POS) in the labels of the source patterns was used to annotate the labels of the target patterns. The empirical results collected during the learning of lexical templates are briefly described below:
• Existing ontologies share the same lexical patterns. For instance, ap-
proximately 60% of the labels that describe an ontological concept make use of an adjective followed by a noun (e.g., spatial region, in- dustrial product, natural hazard, etc.). Other labels use as lexical pattern (≈ 30%) a noun followed by another noun (e.g., transport vehicle, knowledge domain, etc.).
• Ontology labels usually have less than four tokens. Approximately 85%
of labels fulfill this. Thus, for the current prototype we only focus on the definition of lexical templates for compound labels of two o three tokens.
A repository is used to store all the lexical templates obtained for each pair of languages. Table 6.1 shows a sample list of the lexical templates learned to translate compound labels from English into Spanish.
Table 6.1: Some lexical templates to translate a compound label from En- glish into Spanish.
Templates (4/25) Samples of source and target patterns English Spanish
[J1 N2]en→[N2 J1]es spatial region→ regi´on espacial
industrial product→ producto industrial
natural hazard→ peligro natural
[N1 N2]en→[N2⟨pre⟩N1]es transport vehicle→ veh´ıculo de
transporte
knowledge domain→ dominio del
conocimiento research exploration→ exploraci´on de la
investigaci´on [J1 VB2]en→[VB2⟨pre⟩J1]es remote sensing→ detecci´on remota;
detecci´on a distancia [J1N2N3]en→[N2⟨pre⟩N3J1]es associated knowledge dominio de
domain→ conocimiento
asociado J: adjective; N: noun; VB: verb
As an illustrating example of the compositional method, we show in Fig- ure 6.18 the steps of the algorithm when collecting Spanish translations for
the English compound label “AssociateProfessor”, which was introduced in our motivating example. Our system finds ten translations for the token “associate” and one for “professor” (normalized in the first step). In the next step, our tool searches a lexical template (in our repository) to cre- ate candidate translations. In the template found, [J1 N2]en represents the
source pattern in English whilst [N2 J1]es represents the target pattern in
Spanish. In both cases, numbers represent the position of each token of the compound label. Notice that, in the last step the candidate translations “profesor socio” (professor member) and “profesor compa˜nero” (accompa- nying professor) are discarded because they do not fulfill the target pattern.
J : adjective profesor compañero (professor mat)
profesor vinculado (connected profesor) profesor asociado (associate professor) profesor socio (professor member)
[N2 N1]es [N2 J1]es [N2 J1]es [N2 N1]es Target Pattern some Candidate Translations (4/10)
Source Pattern [J1 N2]en
Lexical Template for "associate professor": [J1 N2]en−[N2 J1]es
profesor vinculado (connected professor) profesor asociado (associate professor) compañero (mate)
Step 3: Creating translations into Spanish
profesor (professor)
Step 4: Select candidate translations
asociado (having partial rights and privileges) vinculado (having a logical or causal connection)
some translations of "associate" (4/10)
Step 2: Obtain English−Spanish translations
using Lexical Templates (using Translation service)
translations of "professor" (1/1) AssociateProfessor
associate professor
Step 1: Ontology Label Normalization
N : noun socio (partner)
Figure 6.18: Algorithm to translate the compound label “AssociateProfes- sor” into Spanish.
6.9
Summary of the Chapter
In this chapter we have discussed two important issues related to the on- tology localization activity: life-cycle model and system architecture. We have first described the phases of life-cycle model of the localization activity. A description of the key concepts and elements needed to build a ontology localization system have been included.
Second, we have explained the different modules in the generic architec- ture for an automated ontology localization in collaborative and distributed environments. We have presented the global architecture motivated by the phases identified in the life-cycle model. Also, in order to define the infras- tructure requirements we took different key factors from different software
localization approaches and then we compared them with our own obser- vations in the field. Concretely we described three groups of requirements: collaboration and distribution of the tasks, automated translation, and ex- tensibility.
The Ontology Management Module has been introduced as the mod- ule that enables ontology editors to automatically manage the multilingual content for localization. We have also explained that the control and man- agement of the localization activity is performed by the Localization Man- agement module. The functionality of the Ontology Translator module has been presented later.
Finally, we have presented the LabelTranslator system, which is our approach to automatically localize ontologies among English, Spanish and German. We have included the description of technical details of the main components in the architecture. Furthermore, we have discussed the dif- ferent aspects related to the implementation of the Ontology Repository component included in the Ontology Management module. A description of the capabilities of the implemented interfaces has also been included. With regard to the Localization Management module we have described its main functionalities, which are: synchronize the changes of the ontology and linguistic model and implement the actions described in the collaborative workflow. We have finished with the description of the technical details of the main components in the architecture by commenting on some implemen- tation details of the Ontology Translator module. Here, we have described the translation strategies used to localize simple and compound labels.
Methodological Guidelines
In this chapter we present efficient, prescriptive and detailed methodological guidelines for the ontology localization activity, which are inspired on Soft- ware Engineering methodologies. We first present the scope of the method- ological guidelines together with a brief description of the NeOn Methodol- ogy, which proposes as one of its methodological scenarios the localization of ontologies to support the adaptation of an ontology to different languages and cultures. Second, we describe the process followed to define the ac- tivities and tasks of the methodological guidelines. Finally, we detail the guidelines for the ontology localization activity (including the filling card and the activity workflow).
7.1
Neon Methodology as Framework for the Lo-
calization of Ontological Resources
Contrary to traditional methodologies [Fern´andez-L´opez et al., 1999, Staab et al., 2001, Pinto et al., 2004] that provide methodological guidance for ontology engineering, the NeOn Methodology [Suarez-Figueroa, 2013] iden- tifies a set of flexible scenarios that supports different aspects of the on- tology development process, as well as the reuse and dynamic evolution of networked ontologies in distributed environments, where knowledge is in- troduced by different people (domain experts, ontology practitioners) at different stages of the ontology development process. The nine scenarios proposed in the methodology cover commonly occurring situations in the ontology development process that can be combined according to the on- tology requirements and the existing resources in the domain. Figure 7.1 presents the nine identified scenarios for building ontologies and ontology networks. The directed arrows with numbered circles associated represent the different scenarios. Each scenario covers a specific process or activity (represented with coloured circles or with rounded boxes) that has to be followed to develop an ontology whenever certain requirements or premises
are given.
Figure 7.1: NeOn Methodology scenarios for building ontology networks). The methodological guidelines that we propose in this thesis are intended to assist Scenario 9: Localizing Ontological Resources. The aim is to carry out the localization process of ontologies already conceptualized. Usually, these ontologies are designed without taking into account the multilingual and localization aspects. Therefore, these guidelines aim to reduce costs, improve its quality, and increase the consistency of the localization activity. The proposed guidelines [Espinoza et al., 2012] are particularly intended for ontology stake-holders such as localization managers, translators and reviewers, who are concerned with the ontology localization activity. In ad- dition, they are intended for communities interested in localizing ontologies and those international firms that may promote multilingualism in their working environments for a variety of reasons.
To the best of our knowledge, no guidelines exist for supporting the on- tology localization activity. However, software localization methodologies could be adapted for ontology localization, as these methodologies are very general. In the following section we will define the process followed to de- velop this methodology.