=Paper=
{{Paper
|id=None
|storemode=property
|title=XLore: A Large-scale English-Chinese Bilingual Knowledge Graph
|pdfUrl=https://ceur-ws.org/Vol-1035/iswc2013_demo_31.pdf
|volume=Vol-1035
|dblpUrl=https://dblp.org/rec/conf/semweb/WangLWLLZSLZT13
}}
==XLore: A Large-scale English-Chinese Bilingual Knowledge Graph==
https://ceur-ws.org/Vol-1035/iswc2013_demo_31.pdf
XLore: A Large-scale English-Chinese
Bilingual Knowledge Graph
Zhigang Wang†, Juanzi Li†, Zhichun Wang‡, Shuangjie Li†, Mingyang Li†,
Dongsheng Zhang†, Yao Shi†, Yongbin Liu†, Peng Zhang†, and Jie Tang†
† DCST, Tsinghua University, P.R. China
{wzhigang, ljz, lsj, lmy, zds, sy, lyb, zp,
tangjie}@keg.cs.tsinghua.edu.cn
‡ CIST, Beijing Normal University, P.R. China
zcwang@bnu.edu.cn
Abstract. 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.
1 Introduction
Multilingual knowledge bases are important for the globalization of knowledge
sharing. Knowledge bases such as DBpedia1 , YAGO2 , and BabelNet3 are main-
ly 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 en-
coded in non-English languages is much less than those in English. (ii) The
inconsistency of the large category system in Wikipedia causes incorrect seman-
tic relations between concepts that are defined based on categories. For exam-
ple, “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 heteroge-
nous online wikis to enrich the Chinese knowledge, utilize a classification-based
method to correctly semantify the wikis’ category systems, and employ a cross-
lingual knowledge linking approach to find new cross-lingual links between enti-
ties. Besides, we use a cross-lingual structured knowledge extraction method to
enrich the semantic relations.
1
http://dbpedia.org/
2
http://www.mpi-inf.mpg.de/yago-naga/yago/
3
http://lcl.uniroma1.it/babelnet/
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 Approach
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.
Fig. 1. Overview
Semantifying Online Wikis To semantify the online wikis, we are to pre-
dict 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.
Table 1. Examples of Semantic Relations
Entity 1 Relation Entity 2 Right or Wrong
European Microstates instanceOf Microstates Right
European Microstates instanceOf Europe Wrong
教育人物(Educational Person) subClassOf 人物(Person) Right
教育人物(Educational Person) subClassOf 教育(Education) Wrong
Formally, we learn two series of functions g1 (for English) and g2 (for Chi-
nese) to predict the probabilities to be an is-a relation between two entities.
We define some literal and structural features and train the Logistic Regres-
sion 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 func-
tions 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 inte-
grate the knowledge in different languages, we proposed learning based approach-
es for linking equivalent entities in different languages [1][2]. Several features are
defined based on the link structures in wikis to assess the similarities between
two different entities. Then learning models are trained based on the already
known cross-lingual links in Wikipedia, which afterward predict new equivalen-
t entity pairs. In order to find desired number of new cross-lingual links, we
use concept annotation to enrich the inner links within wikis, which improves
the knowledge linking approach considerably. The knowledge linking process as
a whole can execute iteratively, resulting in large number of new cross-lingual
links.
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 [3]. WikiCiKE is based on
the hypothesis: one can use the rich auxiliary (e.g. English) information to as-
sist the target (e.g. Chinese) infobox extraction. We treat this task as a transfer
learning-based binary classification problem. Given an attribute in the target wi-
ki, WikiCiKE automatically generates the cross-lingual training data and learns
the extractor using TrAdaBoost model. Finally, WikiCiKE uses the learned ex-
tractor to extract the missing value from the unstructured article texts. Our
experiments in [3] demonstrate that WikiCiKE significantly outperforms the
monolingual knowledge extraction method and the translation-based method.
3 XLore System
We construct a unified knowledge graph in the form of RDF and use the Open-
Link Virtuoso server4 for systematical data management. Using the proposed
approach, XLore harvests 856,146 classes , 71,596 properties and 7,854,301 in-
stances 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.
4
http://virtuoso.openlinksw.com/
(a) Number of Linked Concepts (b) Number of Linked Instances
Fig. 2. Statistics of the Linked Entities.
Fig. 3. Interface of XLore System
References
1. Wang, Z., Li, J., Wang, Z., Tang, J.: Cross-lingual knowledge linking across wiki
knowledge bases. WWW’12
2. Wang, Z., Li, J., Tang, J.: Boosting cross-lingual knowledge linking via concept
annotation. IJCAI’13
3. Wang, Z., Li, Z., Li, J., Tang, J., Z.Pan, J.: Transfer learning based cross-lingual
knowledge extraction for wikipedia. ACL’13