Cross-Lingual Infobox Alignment in Wikipedia Using Entity-Attribute Factor Graph

Abstract

Wikipedia infoboxes contain information about article entities in the form of attribute-value pairs, and are thus a very rich source of structured knowledge. However, as the different language versions of Wikipedia evolve independently, it is a promising but challenging problem to find correspondences between infobox attributes in different language editions. In this paper, we propose 8 effective features for cross lingual infobox attribute matching containing categories, templates, attribute labels and values. We propose entity-attribute factor graph to consider not only individual features but also the correlations among attribute pairs. Experiments on the two Wikipedia data sets of English-Chinese and English-French show that proposed approach can achieve high F1-measure: 85.5% and 85.4% respectively on the two data sets. Our proposed approach finds 23,923 new infobox attribute mappings between English and Chinese Wikipedia, and 31,576 between English and French based on no more than six thousand existing matched infobox attributes. We conduct an infobox completion experiment on English-Chinese Wikipedia and complement 76,498 (more than 30% of EN-ZH Wikipedia existing cross-lingual links) pairs of corresponding articles with more than one attribute-value pairs.

1 Introduction

With the rapid evolution of the Internet to be a world-wide global information space, sharing knowledge across different languages becomes an important and challenging task. Cross-lingual knowledge sharing not only benefits knowledge internationalization and globalization, but also has a very wide range of applications such as machine translation [20], information retrieval [19] and multilingual semantic data extraction [7, 9]. Wikipedia is one of the largest multi-lingual encyclopedic knowledge bases on the Web and provides more than 25 million articles in 285 different languages. Therefore, many multilingual knowledge bases (KB) have been constructed based on Wikipedia, such as DBpedia [7], YAGO [9], Bablenet [11] and XLore [18]. Some approaches have been proposed to find cross-lingual links between Wiki articles, e.g., [15,16,17].

There is a large amount of semantic information contained in Wikipedia infoboxes, which provide semi-structured, factual information in the form of attribute-value pairs. Attributes in infoboxes contain valuable semantic information, which play a key role in the construction of a coherent large-scale knowledge base [9]. However, each language version maintains its own set of infoboxes with their own set of attributes, as well as sometimes providing different values for corresponding attributes. Thus, attributes in different Wikipedia must be matched if we want to get coherent knowledge and develop some applications. For instance, inconsistencies among the data provided by different editions for corresponding attributes could be detected automatically. Furthermore, English Wikipedia is obviously larger and of higher quality than low resource languages, which is why we can use attribute alignments to expand and complete infoboxes in other languages, or at least help Wikipedia communities to do so. In addition, the number of existing attribute mappings is limited, e.g., there are more than 100 thousand attributes in English Wikipedia but only about 5 thousand (less than 5%) existing attribute mappings between English and Chinese.

Being aware of the importance of this problem, several approaches have been proposed to find new cross-lingual attribute mappings between Wikis. Bouma et al. [2] found alignments between English and Dutch infobox attributes based on values. Rinser et al. [13] proposed an instance-based schema matching approach to find corresponding attributes between different language infobox templates. Adar et al. [1] defined 26 features, such as equality, word, translation and n-gram features, then applied logistic regression to train a boolean classifier to detect whether two attributes are likely to be equivalent. These methods can be split into two categories: similarity-based and learning-based. Both of them mostly use the information of the attributes themselves and ignore the correlations among attributes within one knowledge base.

Based on our observation, there are several challenges involved in finding multilingual correspondences across infobox attributes. Firstly, there are Polysemy-Attributes (a given attribute can have different semantics, e.g., country can mean nationality of one person or place of a production) and Synonym-Attributes (different attributes can have the same meaning, e.g., alias and other names), which leads to worse performance on label similarity or translation based methods. Secondly, there also exist some problems in the values of attributes: 1. different measurement (e.g., population of Beijing is 21,700,000 in English edition and 2170 ten thousand in Chinese). 2. timeliness (e.g., population of Beijing is 21,150,000 (in 2013) in French edition). In this way, labels and values alone are not credible enough for cross-lingual attribute matching.

In order to solve above problems, we first investigate several effective features considering characteristics of cross-lingual attribute matching problem, and then propose an approach based on factor graph model [6]. The most significant advantage of this model is that it can formalize correlations between attributes explicitly, which is specified in Sect. 3. Specifically, our contributions include:

• We formulate the problem of attribute matching (attribute alignment) across Wikipedia knowledge bases in different language editions, and analyse several effective features based on categories, templates, labels and values.

• We present a supervised method based on an integrated factor graph model, which leverages information from a variety of sources and utilizes the correlations among attribute pairs.

• We conduct experiments to evaluate our approach on existing attribute mappings in the latest Wikipedia. It achieves a high F1-measure 85.5% between English and Chinese and 85.4% between English and French. Furthermore, we run our model on English, Chinese and French Wikipedia, and successfully identify 23,923 new cross-lingual attribute mappings between English and Chinese, 31,576 between English and French.

The rest of this paper is organized as follows, Sect. 2 defines the problem of attribute matching and some related concepts; Sect. 3 describes the proposed approach in detail; Sect. 4 presents the evaluation results; Sect. 5 discusses some related work and finally Sect. 6 concludes this work.

2 Problem Formulation

In this section, we formally define the problem of Wikipedia attribute (property) matching. We define the Wiki knowledge base and elements in it as follows.

Definition 1

Wiki Knowledge Base. We consider each language edition of Wikipedia as a Wiki Knowledge Base, which can be represented as

$$K = \{A,P\}$$

where $A=\{a_i\}^n_{i=1}$ is the set of articles in K and n is the size of A, i.e., the number of articles. $P=\{p_i\}^r_{i=1}$ is the set of attributes in K and r is the size of P.

Fig. 1.

An example of attribute matching

Definition 2

Wiki Article. A Wiki Article can be formally defined as follows,

$$a = (Ti(a),Te(a),Ib(a),C(a))$$

where

• Ti(a) denotes the title of the article a.

• Te(a) denotes the unstructured text description of the article a, in other words, the free-text contents of the article a.

• Ib(a) is the infobox associated with a; specifically, $Ib(a) = \{p_i,val_i\}^k_{i=1}$ represents the list of attribute-value pairs in this article’s infobox, $P(a)=\{p_i\}^k_{i=1}$ represents the set of attributes which appear in Ib of a.

• C(a) denotes the set of categories of the article a.

Definition 3

Attribute. According to the above definitions, an attribute can be defined as a 5-tuple,

$$attr = (L(p),SO(p),AU(p),C(p),T(p))$$

where

• L(p) denotes the label of attribute p.

• $SO(p)=\{(a,val)\mid \forall a\in A, \exists (p,val)\in Ib(a)\}$ denotes a set which contains the subject-object pairs of the attribute. For example, in Fig.  1 , attribute Alma_mater has a pair (Mark Zuckerberg, Harvard University) in $SO (p_{Alma\ mater})$.

• $AU(p)=\{a\mid \forall a, \exists (a,val)\in SO(p)\}$ denotes the set of articles which use attribute p.

• $C(p)=\bigcup \limits _{(p,o)\in Ib(a)} C(a)$ denotes a set of categories in which the attribute appears. For example, C of attribute Born contains a category People.

• $T(p)=\{p_i\}^m_{i=1}$ denotes the Infobox template to which the attribute p belongs.

Definition 4

Attribute Matching (Property Matching). Given two Wiki Knowledge Bases $K_1=\{A_1,P_1\}$ and $K_2=\{A_2,P_2\}$, attribute matching is a process of finding, for each attribute $p_i \in P_1$, one or more equivalent attributes in knowledge base $K_2$. When the two Wiki knowledge bases are in different languages, we call it the cross-lingual attribute matching (infobox alignment) problem. Generally, EL, EC and AL denote the existing cross-lingual links between articles, categories and attributes respectively between different language versions of Wikipedia.

Here, we say two attributes are equivalent if they semantically describe the same type of information about an entity. Figure 1 shows an example of attribute matching results concerning infoboxes of Zuckerberg (CEO of Facebook) in English, Chinese and French Wikipedia.

As shown in Fig. 1, Born, and Naissance are equivalent infobox attributes, which can be easily found according to the values using a translation tool. However, for attribute Net worth and its Chinese corresponding attribute , they have different values because of timeliness, so we cannot find the alignment using value-based method. Furthermore, English Infobox (the left) has an attribute Relatives, which does not exist in other two versions. So we can complete the Chinese infobox of Zuckerberg if we find that is the corresponding attribute of Relatives in Chinese.

3 The Proposed Approach

In this section, we first describe the motivation and overview of our approach, and then we introduce our proposed model in detail.

3.1 Overview

For the problem of Wikipedia attribute matching, existing works [1, 2, 13] mostly used label- and value-based features. Effectiveness of these direct features has been proved. However, as for cross-lingual attribute matching, text similarity cannot be computed directly and machine translation may induce more errors. In this way, only text feature is not enough. There are some works [15, 17] on a similar problem, Wikipedia cross-lingual entity matching, and in these works some useful language-independent features are proposed, such as text hyperlinks. Furthermore, these works also provide large amounts of cross-lingual article links which are very valuable. Inspired by these works, we try to design a model leveraging text, article, category and template features simultaneously. Thus, there are two questions in front of us.

• How to use existing cross-lingual links as seeds to help us find more attribute mappings?

• How to use other information (e.g., article, category and external text) to deal with the lack of information in attribute itself?

3.2 Entity-Attribute Factor Graph Model

Factor graph model [6] has such an assumption that observation data depends on not only local features, but also on relationships with other instances. The characteristic of this model is fit for our problem intuitively, because:

• A pair of attributes is more likely equivalent if they co-occur with aligned attributes in a pair of equivalent articles.

• Template pairs which contain more equivalent attribute pairs tend to be more semantically similar, and other attribute pairs in such templates are more likely equivalent than the ones in other templates.

• Attribute pairs tend to be equivalent if their synonymous pairs are equivalent.

Fig. 2.

An illustration of Entity-Attribute Factor Graph (EAFG) model

In this paper, using definitions in Sect. 2, we formalize the attribute matching problem into a model named Entity-Attribute Factor Graph (EAFG), which is shown in Fig. 2.

Figure 2 contains two parts, the left one is relation graph, which represents several relations in two editions of Wikipedia $K_1$ and $K_2$. Different language versions are separated by a diagonal line. The attribute layer contains the attributes and template relations among them. Similarly, the article layer contains the articles and category relations. The imaginary lines between the two layers denote the relation of usage between articles and attributes, and the red dashed lines denote the existing cross-lingual links. The right one is factor graph, the white nodes are variables, there are two types of variables, $x_i$ and $y_i$. Each candidate pair is mapped to an observed variable $x_i$. The hidden variable $y_i$ represents a Boolean label (equivalent or inequivalent) of the observed variable $x_i$. For example, $x_2$ in Fig. 2 corresponds to a candidate attribute pairs $(p_{i3},p_{j2})$, and there exists a cross-lingual link between $p_{i3}$ and $p_{j2}$, so the hidden variable $y_2$ equals to 1. The black nodes in factor graph are factors, there are three types, f, g, and h. Each type is associated with a kind of feature function which transforms relations into a computable feature.

Now, we define these feature functions in EAFG model in detail:

• Local feature function: $f(y_i,x_i)$ is a feature function which represents the posterior probability of label $y_i$ given $x_i$; it describes local information and similarity on observed variables in EAFG;

• Template feature function: $g(y_i,CO(y_i))$ denotes the correlation between hidden variables according to template information. $CO(y_i)$ is the set of variables having template co-occurrence relation with $y_i$.

• Synonym feature function: $h(y_i,SY(y_i))$ denotes the correlation between hidden variables according to synonymous information. $SY(y_i)$ is the set of variables being semantically equivalent.

According to these feature functions, we can define joint distribution over the Y on our graph model as

$$\begin{aligned} p(Y) = \prod _{i}f(y_i,x_i)g(y_i,CO(y_i))h(y_i,SY(y_i)) \end{aligned}$$

(1)

Then we introduce the definition of three feature functions in detail.

• 1. Local feature function

  $$\begin{aligned} f(y_i,x_i) = \frac{1}{Z_\alpha }\exp \{\alpha ^\mathrm {T}\mathbf {f}(y_i,x_i)\} \end{aligned}$$

  (2)

  where $\mathbf {f}(y_i,x_i)$ = ${<}f_{label}, f_{we}, f_{so}, f_{au}, f_{cate}{>}$ is a vector of features; $\alpha$ denotes the corresponding weights of these features; $x_i$ is a variable corresponding to attribute pair $(p_{a},p_{b})$. Then we describe these five features in detail.

  • (a) Label similarity feature: it computes the Levenshtein distance [3] after translating non-English attribute labels to English ones, and then get the similarity according to it.

    $$\begin{aligned} f_{label} = 1 - \frac{Leven(L(p_{a}),L(p_{b}))}{\max (len(L(p_{a})),len(L(p_{b})))} \end{aligned}$$

    (3)

    where $Leven(L(p_{a}),L(p_{a}))$ denotes the Levenshtein distance between two labels, and len(L(p)) denotes the length of the label of the attribute p.

    Word embedding [10] represents each word as a vector and is able to grasp semantic information. We trained 100-dimension word embeddings on English Wikipedia text and represent each label as a vector (non-English labels are replaced by their translation result). Let WE(p) be the word embedding (a 100-dimension vector) of the label of attribute p, we have

    $$\begin{aligned} f_{we} =1-\frac{ \arccos (\frac{WE(p_a)\cdot WE(p_b)}{\left|| WE(p_a)\right||_2 \times \left||WE(p_b)\right||_2}) }{\pi } \end{aligned}$$

    (4)

    where $\left|| WE(p_a)\right||_2$ denotes the Euclidean norm of the vector $WE(p_a)$, and $f_{we}$ is the cosine similarity between word embeddings of $p_a$ and $p_b$.

  • (b) Subject-object similarity feature: according to Definition 3, we can get a set SO for each attribute and compute the similarity between the two sets. First, we define an equivalence relation between subject-object pairs as

    $$ (s_{i},o_{i}) \equiv (s_{j},o_j) \iff (s_{i},s_{j}) \in EL \wedge o_{i} \equiv o_j$$

    it denotes pair $(s_{i},o_{i})$ in $SO_i$ is equivalent with $(s_{j},o_j)$ in $SO_j$ if and only if there is a cross-lingual link between subjects, and objects are equivalent. The condition of objects being equivalent depends on the data type. For example, for type Integer, the objects should be equal, and for type entity, they should also have a cross-lingual link. $f_{so}$ is defined as

    $$\begin{aligned} f_{so} = \frac{2\times |\{(s_i,o_i)\equiv (s_j,o_j)\mid (s_i,o_i)\in SO(p_{a}),(s_j,o_j)\in SO(p_{b})\}|}{|SO(p_{a})|+|SO(p_{b})|} \end{aligned}$$

    (5)

  • (c) Article-usage feature: according to Definitions 3 and 4, we can define $f_{au}$ as

    $$\begin{aligned} f_{au} = \frac{2\times |\{(a,b)\mid (a,b)\in EL, a\in AU(p_a), b\in AU(p_b)\}|}{|AU(p_a)|+|AU(p_b)|} \end{aligned}$$

    (6)

    this feature represents the similarity between two article sets which contain the two attributes in their infoboxes respectively.

  • (d) Category similarity feature: similarly, we can define $f_{cate}$ as

    $$\begin{aligned} f_{cate} = \frac{2\times |\{(c,c')\mid (c,c')\in EC, c\in C(p_a), c'\in C(p_b)\}|}{|C(p_a)|+|C(p_b)|} \end{aligned}$$

    (7)

    where C(p) is defined in Definition 3 and EC is defined in Definition 4. This feature represents the similarity between two category sets related to the two attributes.

• 2. Template feature function

  $$\begin{aligned} g(y_i,CO(y_i)) = \frac{1}{Z_\beta }\exp \{\sum _{y_j\in CO(y_i)}\beta ^\mathrm {T}\mathbf {g}(y_i,y_j)\} \end{aligned}$$

  (8)

  where $\beta$ denotes the weight remaining to learn, and $\mathbf {g}(y_i,y_j)$ denotes a function to specify whether there exists a template sharing correlation between attribute pairs. Let $(p_{a_i},p_{b_i})$ and $(p_{a_j},p_{b_j})$ be the attribute pairs related with node $y_i$ and $y_j$ respectively in the factor graph. $\mathbf {g}(y_i,y_j) = 1$ if $p_{a_i}$ and $p_{a_j}$ appear in one common template, and so are $p_{b_i}$ and $p_{b_j}$, otherwise 0. It should be noticed that this function is used to capture the relations between candidate attribute mappings.

• 3. Synonym feature function

  $$\begin{aligned} h(y_i,SY(y_i)) = \frac{1}{Z_\gamma }\exp \{\sum _{y_j\in S(y_i)}\gamma ^\mathrm {T}\mathbf {h}(y_i,y_j)\} \end{aligned}$$

  (9)

  where $\gamma$ denotes the weight remaining to learn, and $\mathbf {h}(y_i,y_j)$ denotes the probability of semantically equivalence between $y_i$ and $y_j$. First we define semantic relatedness between two attributes as,

  $$\begin{aligned} SR(p_i,p_j) = \frac{2\times |\{(c_i,c_j)\mid c_i\equiv c_j, c_i\in C(p_i), c_j\in C(p_j)\}|}{|C(p_i)|+|C(p_j)|} \end{aligned}$$

  (10)

  which is similar with Eq. 7, except that $p_i$ and $p_j$ here are from the same language, thus the equivalence between category pairs can be derived directly. Then let $(p_{a_i},p_{b_i})$ and $(p_{a_j},p_{b_j})$ be the attribute pairs related with node $y_i$ and $y_j$ respectively, we have

  $$\begin{aligned} \mathbf {h}(y_i,y_j) = SR(p_{a_i},p_{a_j})\times SR(p_{b_i},p_{b_j}) \end{aligned}$$

  (11)

  Therefore, the purpose of this feature function is to find more cross-lingual attribute mappings using information of synonym within one language edition of data set.

3.3 Learning and Inference

Given a set of labeled nodes (known attribute mappings) in the EAFG, learning the model is to estimate an optimum parameter configuration $\theta =(\alpha , \beta , \gamma )$ to maxmize the log-likelihood function of p(Y). Based on Eqs. 1–11, the joint distribution p(Y) can be denoted as

$$\begin{aligned} p(Y) = \frac{1}{Z}\prod _i\exp \{\theta ^\mathrm {T}(\mathbf {f}(y_i,y_j),\sum _{y_j}\mathbf {g}(y_i,y_j), \sum _{y_j}\mathbf {h}(y_i,y_j))\} \end{aligned}$$

(12)

We use log-likelihood function $\log (p(Y^L))$ as the object function, where $Y^L$ denotes the known labels. Then we apply a gradient descent method to estimate the parameter $\theta$. After learning the optimal parameter $\theta$, we can infer the unknown labels by finding a set of labels which maximizes the joint probability p(Y).

4 Experiments

In this paper, the proposed approach is a general model (translation based features are optional), so we use the data from three language editions of Wikipedia (English, Chinese and French) to evaluate our proposed approach. First we evaluate EAFG model on existing cross-lingual attribute mappings, and then we use our approach to find English-Chinese and English-French mappings within Wikipedia.

4.1 Data Set

We construct two data sets (English-Chinese and English-French) from existing cross-lingual attribute links in Wikipedia. Table 1 shows the size of the 2 data sets. In each data set, we randomly select 2,000 corresponding attribute pairs which are labeled as positive instances. For each positive instance, we generate 5 negative instances by randomly replacing one of the attribute in the pairs with a wrong one.

Table 1. Size of the 2 data sets

4.2 Comparison Methods

We conduct four existing cross-lingual attribute matching methods. They are translation based method Label Matching (LM), Similarity Aggregation (SA) based method, classification based method Support Vector Matching (SVM) and another logistic regression based method (LR-ADAR) on the work of Adar [1]. As for our proposed approach, in order to evaluate the influence of translation tool, we conduct EAFG-NT (No Translation) which is same as EAFG except that it does not use translation-based features.

• Label Matching (LM). This method first uses Google Translation API to translate the labels of attributes in other languages into English, and then matches them. For each attribute pair, they are considered as equivalent attributes if they have strictly the same English labels.

• Similarity Aggregation (SA). This method aggregates several similarities of each attribute pair into a combined one averagely. Here, we compute 5 similarities same as local feature function in Sect. 3, namely label similarity, subject-object similarity, article-usage similarity, category similarity and word embedding similarity.

  $$Sim(p_{i},p_{j}) = \frac{1}{5}(f_{label}+f_{so}+f_{au}+f_{cate}+f_{we})$$

  Then it selects pairs whose similarity is over a threshold $\phi$ as equivalent pairs. In our experiment, we test the parameter $\phi$ from 0.05 to 1.00 increasing by 0.05, and this method achieves the best F1-measure when $\phi =0.75$ on EN-ZH data set, $\phi =0.80$ on EN-FR data set.

• Support Vector Machine (SVM). This method first computes the five similarities in method SA, and then trains a SVM model [4]. Here, we use Scikit-Learn package [12] in our experiment with a linear kernel and parameter $C =1.0$. Finally we predict the equivalence of new attribute pairs using this model. Compared with our approach, this method only uses similarities of attributes as features, and it does not take correlations among these instances into consideration.

• Logistic Regression (LR-ADAR) In [1], the author defined 26 features and trained a logistic regression model to solve this problem. They obtained good results in their experiments, so we implement this method as a comparison. Here we also use Scikit-Learn package to train a logistic regression model with 17 of their features (removing some language features because they are not suitable for Chinese). In our experiment, it achieves the best result when we use parameter $C =10$ and L1-regularization.

4.3 Performance Metrics

We use Precision, Recall and F1-measure to evaluate different attribute matching methods. They are defined as usual: Precision is the percentage of correct discovered matched in all discovered matches; Recall is the percentage of correct discovered matches in all correct matches; F1-Measure is the harmonic mean of precision and recall. The data sets we use are described in Sect. 4.1.

4.4 Settings

For SVM, LR-ADAR and EAFG, we conduct 10-fold cross validation on the evaluation data set. EAFG uses 0.001 learning rate and runs 1000 iterations in all the experiments, and SVM and LR-ADAR runs with settings described in the above. As mentioned before, translation tool is optional in our approach, so we also implement method EAFG-NT for comparison. All experiments are carried out on a Ubuntu 14.04 server with 2.8 GHz CPU (8 cores) and 128 GB memory.

Table 2. Perfomance of 5 methods on English-Chinese and English-French data sets.

4.5 Results Analysis

Table 2 shows the performance of these 5 methods on English-Chinese (EN-ZH) and English-French (EN-FR) data sets. For EN-ZH data set, according to the results, the LM method gets a high precision of 97.3%, but its recall is only 26.1%. Apparently, the variety of translation results and too strict matching condition are the main reasons of the result. By using similarities on various information, SA improves recall significantly in comparison to LM, but it does not achieve good precision because averaging strategy is too simple. SVM and LR-ADAR are both learning-based methods. SVM method gets a precision of 87.5% with a recall 75.2%. Compared with SVM, LR-ADAR gets better precision but lower recall, and outperforms SVM by 1.0% in terms of F1-measure. Our method EAFG uses the same training data with SVM, and outperforms SVM by 4.6% in terms of F1-measure. EAFG get similar precision with LR-ADAR, but EAFG is able to discover more attribute mappings by considering the correlation between attribute pairs. EAFG-NT only uses language-independent features, although it does not work as well as EAFG, it still outperforms SVM by 0.5%, which indicates that correlations among attributes are helpful for the problem indeed. As for EN-FR data set, most of these methods get better precision than EN-ZH, and we think it is because English and French are both European languages. Correspondingly, we can get similar conclusions from the experiment on EN-FR data set.

4.6 Discovering New Cross-Lingual Attribute Mappings in Wikipedia

The motivation of this work is to find more attribute mappings among different language versions of Wikipedia. Therefore, we applied our proposed EAFG to align attributes in English, Chinese and French Wikipedia. First, we extract 107,302, 56,140 and 85,841 attributes from English, Chinese and French Wikipedia respectively. The existing attribute mappings are used for training, and the learned model is employed to predict the correspondence between cross-lingual attribute pairs. Both training and prediction are completed on a server with a 2.8 GHz CPU (32 cores) and 384 GB memory, and it costs 13 h and 21 h for EN-ZH and EN-FR data set respectively. Finally we get 23,923 new attribute mappings between English and Chinese Wikipedia, and 31,576 mappings between English and French. Table 3 presents a few examples of the discovered mappings.

Table 3. Examples of discovered attribute mappings

4.7 Wikipedia Infobox Completion

Apparently, we can transfer infobox information that is missing in one language from other languages in which the information is already present, if we have the alignment of attributes. In this paper, we try to complement Chinese and English Wikipedia infoboxes from each other using the attribute alignments obtained above EAFG. Firstly, we extract 223,159 existing corresponding English-Chinese article pairs, and finally 76,498 article pairs are replenished by at least 1 attribute value. Figure 3 shows the number of added attribute values with respect to each article. The maximum number of added attribute values for one article is 34 and the average is 5.75, which indicates that infoboxes in Chinese and English both benefit a lot from the attribute alignments.

Fig. 3.

Statistics of EN-ZH infobox complementing

We also count the times of each attribute being added into Chinese infoboxes, and list the top 20 attributes in Fig. 4. It should be noticed that most of the attributes are from these categories: Person (e.g., Born and Nationality), Location (e.g., time-zone and Original language) and Film (e.g., Director and Producer). The reason is that entities of these categories tend to have strong local features, and thus lead to imbalance of information among different language versions of Wikipedia. For example, a recent TV play The Journey of Flower ¹ ( ² in Chinese) is very popular in China and its Chinese Wikipedia page contains elaborate information. In this experiment, we add 7 attribute values (such as (editor, Tianen Su), (original channel, Hunan Satellite)) from Chinese to English Wikipedia with respect to this entity (i.e., The Journey of Flower).

Fig. 4.

Top 20 Attributes in EN-ZH infobox complementing

5 Related Work

In this section, we review some related work.

5.1 Wikipedia Infobox Alignment

Though there have been some works on Wikipedia cross-lingual infobox alignment (attribute matching) and its applications in the real world. Adar [1] used a supervised classifier to identify cross-language infobox alignments. They use 26 features, including equality and n-gram to train the classifier. Through a 10-fold cross-validation experiment on English, German, French and Spanish, they report having achieved 90.7% accuracy. Bouma [2] proposed a value-based method for matching infobox attributes. They first normalized all infobox attribute values, such as numbers, data formats and some units, and then matched the attributes according to the equality between English and Dutch Wikipedia. Rinser [13] proposed an instance-based attribute matching approach. They first matched entities in different language editions of Wikipedia, then they compared the values in attribute pairs and got final results using the entity mappings. However, these works did not consider the correlations among candidate attribute pairs, which is proved to be effective for attribute matching in our work.

5.2 Ontology Schema Matching

Ontology schema matching [14] is another related problem which mainly aims to get alignments of concepts and properties. Currently, some works focus on monolingual matching tasks, such as SOCOM [5] and RiMOM [8, 21]. These systems deal with the cross-lingual ontology matching problem mainly using machine translation tools to bridge the gap between languages. In our approach, translation-based features are optional.

6 Conclusions and Future Work

In this paper, we propose a cross-lingual attribute matching approach. Our approach integrates several feature functions in a factor graph model (EAFG), including labels, templates, categories and attribute correlations to predict new cross-lingual attribute mappings. Evaluations on existing mappings show that our approach can achieve high F1-measure with 85.5% and 85.4% on English-Chinese and English-French Wikipedia respectively. Using our approach, we have found 23,923 new attribute mappings between English and Chinese Wikipedia and 31,576 between English and French. It is obvious that article and attribute mappings can benefit each other. Therefore, in the future, we are going to design a framework which can simultaneously and iteratively align all of the elements in Wikipedia.