Nowadays, there are many natural languages on the Web, and we can expect that they will stay there even with the development of the Semantic Web. Though the RDF model enables structuring information in a uni ed way, the resources can be described using di erent natural languages. To nd information about the same resource across di erent languages, we need to link identical resources together. In this paper we present an instance-based approach for resource interlinking. We also show how a problem of graph matching can be converted into a document matching for discovering cross-lingual mappings across RDF data sets.
Due to the Resource Description Framework (RDF), the information on the Web can be turned from the unstructured mass into the structured data represented in the form of triples. The Linked Open Data (LOD) cloud containing billions of triples is constantly growing. Since data sets are created independently, there can be several Uniform Resource Identi ers (URIs) denoting the same entity across di erent RDF data sets. As a result, one needs to address the problem of entity resolution: identify and interlink the same entity across multiple data sources.
The RDF syntax is relatively simple and unambiguous: RDF = graph + identi ers (labels). This is what the identi cation of resources can be based on. However, this problem can become particularly di cult when there are multilingual elements in a graph as a simple string matching technique is doomed to fail. Hence, speci c Natural Language Processing (NLP) techniques must be considered.
Our research problem is to nd out methods for linking the same resource located in several RDF data sets and described in various natural languages and study the impact of available NLP techniques on the interlinking procedure.
Internet is a multilingual system, and we believe that it will continue to accommodate a diversity of natural languages despite the development of the Semantic Web. Even though there are many resources in English, some other languages occupy a decent portion of the Web space as well (see Internet world users by language statistics 1). And we expect the necessity to tackle the multilinguality problem to persist. There are many resources that could be interlinked. At present, the number of languages 2 of RDF data sets amounts to 503.
The importance of cross-lingual mappings has been discussed in several works [1{3].
Recently a Best Practices for Multilingual Linked Open Data Community Group 3 has been created to elaborate a large spectrum of practices with regard to multilingual LOD.
Availability of the cross-lingual links is imperative for several neighboring
research areas. For example, to overcome the problem of ontology
heterogeneity, some research has been done on monolingual ontology integration based on
instances interlinked by owl:sameAs [
The owl:sameAs links between instances can be also valuable in other
applications such as Question Answering over multilingual structured knowledge-base
[
Thus, the growing number of data sources in RDF format with multilingual labels and the importance of cross-lingual links for other Semantic Web applications motivate our interest in cross-lingual link discovery. 3
The problem of searching for the same entity across multiple sources has been
investigated in several research elds. In database community, it is known as
instance identi cation, record linkage or record matching problem. In [
In the eld of Information Retrieval (IR), within the framework of the CrossLanguage Evaluation Forum (CLEF)4, the Web People Search Evaluation Campaigns (2007-2010)5 focused on the Web People Search and person name ambiguity on Web pages and aimed at building a system which could estimate the
2 http://stats.lod2.eu/languages 3 http://www.w3.org/community/bpmlod/ 4 http://www.clef-initiative.eu/ 5 http://nlp.uned.es/weps/weps-3 number of referents and cluster Web pages that refer to the same individual into one group. The research was performed on monolingual data.
Cross-lingual entity linking has been addressed in Knowledge Base
Population track (KBP2011)[
In contrast to the research outlined above, we aim at providing insights into the problem of cross-lingual interlinking from the point where data are already in RDF format, and we can vary di erent parameters in order to determine their impact on the interlinking operation.
In the Semantic Web, interlinking resources that represent the same
realworld object and that are scattered across multiple Linked Data sets is a widely
researched topic. Within the Data Interlinking track (IM@OAEI 2011), several
interlinking systems have been proposed [9{13]. All of the systems were
evaluated on monolingual data sets. Recent developments have been made also in
multilingual ontology matching [
To the best of our knowledge, there is no interlinking system speci cally designed to link RDF data sets with multilingual labels. 4
The goal of our work is to provide methods to link interrelated resources across
multilingual RDF data sets. For now, we restrict ourselves to owl:sameAs link
[
Our general research question is: To what extent is it possible to interlink data sets in di erent languages? To answer this question, within the framework that we describe in the Proposed Approach section, we need to explore which parameters in uence this task. More speci cally: 1. How to represent entities from RDF graphs? { What is the optimal distance for collecting language elements in traversal? { Is it necessary to preserve the structure of the graph in a virtual document by weighting the path length? 2. How to make entities described in di erent natural languages comparable? { What are the most appropriate Machine Translation techniques (rulebased, statistical, hybrid)? { What is the impact of translating one language into another or pivot language? { How does the output of similarity measures vary according to the context? All these parameters will be studied with respect to speci c contexts (language pairs, data set types, amount of textual data available). We also plan to experiment with graph matching techniques to see the di erence with a translationbased approach. Apart from Machine Translation, we will explore techniques used for word alignment, thesaurus-based word sense disambiguation, multilingual document ranking, and mapping to multilingual lexicons. 5
We introduce several hypotheses that we would like to test in our research. 1. If two URIs denote the same real-world object, the descriptions of the properties of this object should overlap with each other. 2. If descriptions are in di erent natural languages, then NLP techniques could help to decrease uncertainty across a set of resources. 3. If the descriptions of an entity overlap signi cantly, the similarity between them will be higher than between other entities. 4. If the degree of similarity depends on the available language context for each entity, then the more language data there are, the better will be the matching results. 5. If language data can be taken from two sources in RDF graph: property names and literals; then literals are more important since they are more informative. 6
Due to the presence of natural language terms in RDF graphs, we adopt a language-oriented approach.
The proposed approach includes several steps (see Figure 1). 1. Given two data sets with a resource representation in di erent natural language, extract language data for each URI. Thus, we create a "virtual" document for each URI. 2. Compare virtual documents in pairs from both sets. 3. Find the maximum similarity between two representations of the resource. 4. Establish an owl:sameAs link between the two most similar representations.
One should mind the following aspects of this approach: DOCUMENT(ru) 2 1 DOCUMENT(en)
SIMILARITY 3
DOCUMENT(en) translation
DOCUMENT(zh)
{ The idea of creating a "virtual" document has been employed in ontology
matching [
There are some complications as to textual data. Two scenarios are possible: { URI can be looked up and the textual data extracted (as in case of DBpedia) { No extra textual data are available per URI except the data in a graph itself.
To overcome this lack of context for a particular resource, we propose to browse a graph up to n+1 hops from the URI under investigation and collect data along the way. The data carriers are property names and literals. Thus, a virtual document for a particular URI will be the accumulation of data gathered during graph traversal.
This way of collecting a "pro le" for a resource entails a question: does the di erence of two graph structures a ect the results of interlinking? On the one hand, taking into account the success of statistical machine translation based on statistical modelling and probabilities, the order of words is not always that important. It would be interesting to see whether it holds for RDF interlinking as well. On the other hand, we can try to preserve the order of collected properties and literals in a virtual document by putting weight for each language element. The further it is from the URI at question, the lower the weight. Term weight can be assigned by computing termf requency in a document or distribution of terms across a collection of documents known as inverse documentf requency (IDF). Terms that appear in few documents can be discriminative with regard to the rest of the documents. Combination of both TF x IDF is widely used in vector space models.
Once virtual documents are collected from both graphs, the documents will be compared and results evaluated. 7
We believe that we can succeed in nding the solution for our research topic because we plan to put our research on a solid foundation and combine different methods to achieve the task. In traditional Web, there has been much work done on multilingual NLP, i.e. language identi cation, machine translation, cross-language information retrieval. We are going to conduct series of experiments and see what works and how we can improve what does not work. This would allow us to preserve only the best practices and nally crystallize a solution to the problem. The author of this research proposal is also guided by the specialists in the domain that will contribute to the right choice of the research direction. 8
Evaluation means comparing the retrieved links against some reference. Standard
measures usually serve for evaluation of an interlinking system (Precision, Recall,
F-measure). The biggest challenge for evaluating our system is the absence of
standard benchmark tests. As described in [