Current Wikipedia-based multilingual knowledge bases still suffer the following problems: (i) the scarcity of non-English knowledge, (ii) the noise in the semantic relations and (iii) the limited coverage of equivalent cross-lingual entities. In this demo, we present a large-scale bilingual knowledge graph named XLore, which has adequately solved the above problems.
Multilingual knowledge bases are important for the globalization of knowledge sharing. Knowledge bases such as DBpedia1, YAGO2, and BabelNet3 are mainly built upon the multilingual Wikipedia. Some problems are to be addressed: (i) The imbalanced sizes of different Wikipedia language versions lead to the highly imbalanced knowledge distribution in different languages. Knowledge encoded in non-English languages is much less than those in English. (ii) The inconsistency of the large category system in Wikipedia causes incorrect semantic relations between concepts that are defined based on categories. For example, “Wikipedia-books-on-people is the subCategoryOf People” will lead to the wrong “Wikipedia-books-on-people is subClassOf People” in DBpedia’s SKOS schema. (iii) Integrated by directly using cross-lingual links in Wikipedia, the amount of integrated multilingual knowledge totally depends on these existing cross-lingual links.
In this demo, we present an English-Chinese bilingual knowledge graph named XLore to adequately solve the above problems. We use much larger heterogenous online wikis to enrich the Chinese knowledge, utilize a classification-based method to correctly semantify the wikis’ category systems, and employ a crosslingual knowledge linking approach to find new cross-lingual links between entities. Besides, we use a cross-lingual structured knowledge extraction method to enrich the semantic relations.
To the best of our knowledge, XLore is the first large-scale cross-lingual knowledge graph with balanced amount of Chinese-English knowledge. XLore gives a new way for building such a knowledge graph across any two languages. 2
As shown in Figure 1, the building of XLore contains three stages: (1) Data Preprocessing: First we collect and clean the data sets from four online wikis, namely English Wikipedia, Chinese Wikipedia, Baidu Baike and Hudong Baike. (2) Knowledge Graph Building : Next, we learn the cross-lingual ontology as follows: semantify the online wikis to predict correct semantic relations, conduct cross-lingual knowledge linking to integrate the heterogenous wikis together, and extract the structured knowledge to enrich more relations in the graph. (3) Knowledge Query: Finally, we construct an online system for knowledge acquisition.
Semantifying Online Wikis To semantify the online wikis, we are to predict the correct subClassOf and instanceOf relations between two entities. We view both the correct subClassOf and the instanceOf relations as is-a relations. Table 1 shows some examples about the semantic relations.
Entity 1 Relation Entity 2 European Microstates instanceOf Microstates
European Microstates instanceOf Europe 教育人物(Educational Person) subClassOf 人物(Person) 教育人物(Educational Person) subClassOf 教育(Education) Right or Wrong
Right Wrong Right Wrong
Formally, we learn two series of functions g1 (for English) and g2 (for Chinese) to predict the probabilities to be an is-a relation between two entities. We define some literal and structural features and train the Logistic Regression models. The most important features are the head words’ singular/plural forms of English entities and the substring relationship between the labels of Chinese entities. By iteratively expanding the training data sets, both the functions achieve over 90.48% F1-score. To keep the semantic relatedness, we treat the incorrect relations as the subTopicOf relations and import these relations into the RDF database too.
Cross-lingual Knowledge Linking via Concept Annotation To
integrate the knowledge in different languages, we proposed learning based
approaches for linking equivalent entities in different languages [
Cross-lingual Structured Knowledge Extraction To enrich kinds of
relations in XLore, we apply our cross-lingual knowledge extraction framework
named WikiCiKE to complete the missing infoboxes [
We construct a unified knowledge graph in the form of RDF and use the OpenLink Virtuoso server4 for systematical data management. Using the proposed approach, XLore harvests 856,146 classes , 71,596 properties and 7,854,301 instances across English and Chinese. Figure 2 gives a brief statistics the number of linked entities from different online wikis.
We also deploy an online system to illustrate our XLore. As shown in Figure 3, the system supports the keyword-based or SPARQL queries, gives the statistical information, offers visualization demonstrations, etc. A live demonstration of the system can be found at http://www.youtube.com/watch?v=QKA-RYFfztA. We invite the readers to try our XLore prototype at http://xlore.org.
(a) Number of Linked Concepts (b) Number of Linked Instances