The paper presents a novel technique for aligning cross-lingual ontologies that does not rely on machine translation, but uses the large multilingual semantic network BabelNet as a source of background knowledge. In addition, our approach applies a novel orchestration of the components of the matching workflow. We demonstrate that our method outperforms considerably the best techniques in the state-of-the-art.
In spite of the considerable advance that has been made in the field of ontology matching recently, many questions remain open [1]. The current work addresses the challenge of using background knowledge with a focus on aligning cross-lingual ontologies, i.e., defined in different natural languages [2].
Indeed, considering multilingual and cross-lingual information is becoming more and more important, in view particularly of the growing number of contentcreating non-English users and the clear demand of cross-language interoperability. In the context of the web of data, it is important to propose procedures for linking vocabularies across natural languages, in order to foster the creation of a veritable global information network.
The use of different natural languages in the concepts and relations labeling process is becoming an important source of ontology heterogeneity. The methods that have been proposed to deal with it most commonly rely on automatic translation of labels to a single target language [3] or apply machine learning techniques [2]. However, machine translation tolerates low precision levels and machine learning methods require large training corpus that is rarely available in an ontology matching scenario. An inherent problem of translation is that there is often a lack of exact one-to-one correspondence between the terms in different natural languages.
We present LYAM++ (Yet Another Matcher -Light), a fully automatic crosslingual ontology matching system that does not rely on machine translation. Instead, we make use of the openly available general-purpose multilingual semantic network BabelNet 1 in order to recreate the missing semantic context in Fig. 1: The processing pipeline of LYAM++. the matching process. Another original feature of our approach is the choice of orchestration of the matching workflow. Our experiments on the MultiFarm2 benchmark data show that (1) our method outperforms the best approaches in the current state-of-the-art and (2) the novel workflow orchestration provides better results compared to the classical one.
The workflow of LYAM++ is given in Fig. 1. We take as an input a source ontology S, given in a natural language l S and a target ontology T , given in a language l T . The overall processes consists of four main components: a terminological multilingual matcher, a mapping selection module and, finally, a structural matcher. One of the original contributions of this work is the choice of orchestration of these components. Indeed, the places of the mapping selection module and the structural matcher are reversed in the existing OM tools [4]. However, we wanted to ensure that we feed only good quality mappings to the structural matcher, therefore we decided to filter the discovered correspondences right after producing the initial alignment. This decision is supported experimentally in the following section.
The multilingual terminological matching module, the second contribution described in this paper, acts on the one hand as a preprocessing component and, on the other hand -as a light-weight terminological matcher between crosslingual labels. We start by splitting the elements of each ontology in three groups: labels of classes, labels of object properties and labels of data object properties (in colors blue, black and red in the figure), since these groups of elements are to be aligned separately. A standard preprocessing procedure is applied on these sets of labels, comprising character normalization, stop-words filtering, tokenization and lemmatization. The tokens of the elements of T are then aligned to BabelNet. At first, every token of a given label s in S is enriched by related terms and synonyms from BabelNet and all of these terms are represented in the language l T , which makes these terms comparable to the tokens of the labels in T . A simple similarity evaluation by the help of the Jaccard coefficient selects the term in each set of related terms corresponding to a given token from s that has the highest score with respect to every token in each label of T . This helps to restitute the label s in the language l T . Finally, the labels in each group of S and T , seen as sets of tokens, are compared by using the Soft TFIDF similarity measure [5], which produces an intermediate terminological alignment.
The three remaining components are standard OM modules [4], although ordered in a new manner. The Mapping selection is a module that transforms the initial 1 to many mapping to a 1:1 alignment based on the principle of iteratively retaining the pairs of concepts with maximal value of similarity. Finally, the structural matcher component filters the trustworthy pairs of aligned concepts by looking at the similarity values produced for their parents and their children in the ontology hierarchies.
The system is not yet available online. The reason for that is that it depends heavily on the use of BabelNet, which is a protected source. We are working on implementing a sharable version of LYAM++ making use of different open access background knowledge sources.
The alignments produced by LYAM++ for this year's Multifarm track can be found under the following link: http://www.lirmm.fr/benellefi/Lyam++.rar
We have evaluated our approach on data coming from the ontology alignment evaluation initiative (OAEI) 3 and particularly Multifarm-a benchmark designed for evaluating cross-lingual ontology matching systems. Multifarm data consist of a set of 7 ontologies originally coming from the Conference benchmark of OAEI, translated into 8 languages. Two evaluation tasks are defined: task 1 consists in matching two different ontologies given in different languages, while task 2 aims to align different language versions of one single ontology.
We have performed experiments on both tasks by using the pairs of languages given in the summary of our results in Table 1.
In another experiment, we have evaluated the results obtained by using our novel orchestration of matching components, as compared to the standard orchestration. The figures in Table 2 show that the workflow proposed in this paper acts in favor of achieving better results as compared to the standard method. 3 Discussions on the way to improve the proposed system Currently, we are working on enhancing the system in order to make applicable to the general ontology matching problem and not only to cross-lingual ones. We have generated first results on the Conference benchmark without any modification in the settings and our results are quite promising. For the majority of the datasets (ontology pairs) our system achieves a f-score almost as good as the f-score of AML, the best performing system on that track.
We consider that a key feature for the improvement of our system is the appropriate choice of background knowledge. In order to improve the results achieved on the Conference track, we plan to use monolingual general purpose background knowledge (for example, the english subgraphs of YAGO or DBPedia) instead of BabelNet.
We intend to use domain specific background knowledge in order to solve alignment problems in specific areas of knowledge. More precisely, we plan to participate on the Anatomy track by testing different kinds of domain specific background knowledge, such as UMLS or other.
We presented an efficient approach for aligning cross-lingual ontologies by using the multilingual lexical database BabelNet. Subjects of ongoing and future work are (1) testing and evaluating different sources of external knowledge, (2) applying the approach to a larger set of languages and (3) adaptation of the approach to the monolingual case and studying the use of background knowledge in a monolingual ontology matching scenario.