Wikipedia has emerged as the largest multilingual, web based general reference work on the Internet. A huge amount of human resources have been invested in the creation and update of Wikipedia articles which are ideally complemented by so-called infobox templates de ning the type of the underlying article. It has been observed that the Wikipedia infobox type information is often incomplete and inconsistent due to various reasons. However, the Wikipedia infobox type information plays a fundamental role for the RDF type information of Wikipedia based Knowledge Graphs such as DBpedia. This stimulates the need of always having the correct and complete infobox type information. In this work, we propose an approach to predict Wikipedia infobox types by using word embeddings on categories of Wikipedia articles, and analyze the impact of using minimal information from the Wikipedia articles in the prediction process.
Wikipedia has become the most widely used and largest multilingual open encyclopedia. Huge amount of human skills, expertise and e orts goes in for the creation of Wikipedia articles. It comprises of both structured and unstructured or free text. Structured data in Wikipedia is represented in the form of an infobox containing property value pairs summarizing the information content of the article. An infobox is a xed-format table usually added to consistently present a summary of some unifying aspects that the articles share and sometimes to improve navigation to other interrelated articles. Furthermore, infobox information is widely used in di erent Knowledge Graphs (KGs) such as DBpedia.
Wikipedia infobox templates are created and assigned based on the
categorical type of the article, i.e. articles belonging to a speci c genre or type should be
assigned the same template. The assignment of the infobox type to a Wikipedia
article is executed based on the discussions between the contributors and the
editors of the content of the Wikipedia article. However, no integrity tests are
conducted to determine the correctness of the infobox assignment. This leads to
the assignment of incomplete and or incorrect infobox type information which
eventually leads to erroneous RDF type information in the KGs [
The prediction of Wikipedia infobox type can be viewed upon as a text
classi cation problem [
Wu et al.[
The work presented in this paper is inspired by another work on Wikipedia
infobox type prediction presented by Biswas et al.[
In this work, the infobox type prediction problem is considered as a text classication problem, in which infobox types are regarded as classes.
Wikipedia Infobox Type Prediction Features. Three features from the Wikipedia articles are being used for the classi cation process.
{ Table of Contents (TOC) is the collection of section headers and sub headers of the Wikipedia articles. { Abstract (A) of the Wikipedia articles i.e. summary of the entire content. { Categories (C) is a list of Wikipedia's main categorization system, intended to group together articles on similar subjects.
Feature Vector. Word2Vec [
In this paper, the Google pre-trained word vectors 3 of length 300 are used to generate word vectors for each word present in the TOC, abstract as well as for the categories. The Google pre-trained word2vec model includes word vectors for a vocabulary of three million words and phrases trained on roughly 100 billion words from a Google News dataset.
For each Wikipedia articlean abstract vector, a TOC vector and a category vector are generated by performing vector addition on all the word vectors of the abstract and normalized by the total number of words present in each of these features. Finally, a document vector is generated by concatenating these three vectors as shown in Figure 1.
Classi cation. Two classi ers have been trained to predict the Wikipedia
infobox types. The aforementioned document vector is used as the feature vector
in the classi cation method using a Random Forest(RF) Classi er. For a
multilabel convolutional neural network(CNN), categories and TOC are considered
as free text and sentence classi cation [
The classi ers have been trained on the most popular 30 infobox types with 5000 articles for each type from the Wikipedia 2016 version. Features generated using TF-IDF have been used as a baseline.
The experiments established the fact that categories of Wikipedia articles play a vital role to determine the infobox type. With CNN, categories can predict the infobox types with a micro F1-score of 96.8% which is 0.7% better than our previous results obtained when the prediction was based on the entire abstract and TOC combined as shown in Table 1. Furthermore, the word embeddings approach performs much better than the TF-IDF baseline, since word embeddings are able to capture semantic similarities. For instance, the Wikipedia article of the album The Wall by Pink Floyd is assigned to most of the categories containing the word album. Furthermore, using categories only results in higher scores than using the entire abstract. Moreover, the prediction 3 https://code.google.com/archive/p/word2vec/
Feature Set
With Embedding
TF-IDF RF(CV) RF(Split) CNN
RF(CV) RF(Split) results are slightly better if categories are considered together with the rst sentence of the abstract instead of considering the whole abstract, which means that less information is su cient to infer Wikipedia infoboxes. Furthermore, in all the experiments word embeddings perform better than TF-IDF. However, rather similar results are obtained when TOC and categories are combined with the entire abstract as well as only with the rst sentence of the abstract. 5
In this paper, the achieved results strengthen the fact that Wikipedia categories as well as minimal text plays a vital role in the prediction of infobox types. In future we intend to design a semi-supervised approach to correct existing Wikipedia infobox types and to predict infobox types for newly created articles.