Comparable corpora can be used for many linguistic tasks such as bilingual lexicon extraction. By improving the quality of comparable corpora, we improve the quality of the extraction. This article describes some strategies to build comparable corpora from Wikipedia and proposes a measure of comparability. Experiments were performed on Portuguese, Spanish, and English Wikipedia.
Wikipedia is a free, multilingual, and
collaborative encyclopedia containing entries
(called “articles”) for almost 300 languages
(281 in July 2011). English is the more
representative one with about 3 million
articles. However, Wikipedia is not a parallel
corpus as their articles are not translations
from one language into another. Many works
have been published in the last years
focused on its use and exploitation for
multilingual tasks in natural language processing:
extraction of bilingual dictionaries
In addition, multilingual articles of
Wikipedia have been used as a source to build
comparable corpora (Gamallo y Gonz´alez,
2010). The EAGLES - Expert Advisory
Group on Language Engineering Standards
Guidelines (see http://www.ilc.pi.cnr.
it/EAGLES96/browse.html) defines a
“comparable corpus” as one which selects
similar texts in more than one language or
variety. One of the main advantages of
comparable corpora is their versatility to be used
in many linguistic tasks
By taking into account multilingual
potentialities of Wikipedia, our main objective
is to define a method to measure the
similarity (or degree of comparability) of
different comparable corpora built from
Wikipedia. For this purpose, first we describe some
strategies to extract monolingual corpora in
Portuguese, Spanish, and English from
Wikipedia, by making use of some categories
(“Archaeology”, “Biology”, “Physics”, etc.)
to make them comparable according to a
specific topic. These strategies were
described in detail in (Gamallo y Gonz´alez, 2010).
Then, we propose a measure of
comparability to verify whether the corpora are lowly
or highly comparable. For many extraction
tasks, such as bilingual lexicon extraction,
using highly comparable corpora often leads
to better results. There are some works
proposing comparability measures between
monolingual corpora
This paper is organized as follows. Section 2 introduces two strategies to build comparable corpora from Wikipedia. Next, in Section 3, we propose some comparability measures. Then, Section 4 describe some experiments performed in order to measure the comparability between different corpora built using the strategies defined in Sec. 2 . The last section discusses future tasks that will be implemented in order to extend and improve our tools. 2.
Two strategies to Build Wikipedia-Based Comparable Corpora
The input of our strategies is CorpusPedia1, a friendly and easy-to-use XML structure, generated from Wikipedia dump files. In CorpusPedia, all the internal links found in the text are put in a vocabulary list identified with the tag links. In the same way, all the categories (or topics) used to classify each article are inserted in the tag category. In addition, there is a tag called translations which codifies a list of interlanguage links (i.e., links to the same articles in other languages) found in each article. Categories and translations are very useful features to build comparable corpora. Given these features, we developed two strategies aimed to extract corpora with different degrees of comparability.
tracts those articles in two languages having in common the same topic, where the topic is represented by a category and its translation (for instance, the English-Spanish pair “ArchaeologyArqueolog´ıa”). It results in a not-aligned comparable corpus, consisting of texts in two languages. We called it “notaligned” because the version of an article in one language may have not its corresponding version in the other language. Aligned Corpus The goal is to extract pairs of bilingual articles related by interlanguage links if, at least, one of both contains a required category. It results in a comparable corpus that is aligned article by article.
In Section 4, we will measure the degree of comparability of corpora built by means of these two strategies. Before that, we will define how to measure comparability between Wikipedia-based corpora.
Comparability Measures
For a comparable corpus C of Wikipedia articles, constituted for instance by a Portuguese part Cp and a Spanish part Cs, a comparability coefficient can be defined on the basis 1The software to build CorpusPedia, as well as CorpusPedia files for English, French, Spanish, Portuguese, and Galician, are freely available at http: //gramatica.usc.es/pln/ of finding, for each Portuguese term tp in the vocabulary Cpv of Cp, its interlanguage link (or translation) in the vocabulary Csv of Cs. The vocabulary of a Wikipedia corpus is the set of “internal links” found in that corpus. So, the two corpus parts, Cp and Cs, tend to have a high degree of comparability if we find many internal links in Cpv that can be translated (by means of interlanguage links) into many internal links in Csv. Let T ransbin(tp, Csv) be a binary function which returns 1 if the translation of the Portuguese term tp is found in the Spanish vocabulary Csv. The binary Dice coefficient, Dicebin, between two parts of a comparable corpus C is then defined as: Dicebin(Cp, Cs) = 2 Ptp∈Cpv T ransbin(tp, Csv)
|Cpv| + |Csv|
We consider that it is not necessary to define the counterpart of the translation function, since the number of ambiguous terms is very low in Wikipedia, and most cases of ambiguity are solved with the so-called “disambiguated pages”.
To avoid a bias towards common internal links, that is, towards those links occurring in most articles, we define a specific version of tf idf weight for each term. In particular, tf idf (tp) is the frequency of term tp in the Portuguese part of the comparable corpus, multiplied by its inverse article frequency in the whole Portuguese Wikipedia. By taking into account the tf idf of terms, we can define a weighted measure of comparability. Let T ranstf idf (tp, Csv) be a function which returns the smallest value (min) of two tf idf scores, both tf idf (tp) and tf idf (ts), where ts is the Spanish translation of tp in the Spanish part Cs. The weighted Dice coefficient, Dicetf idf , between two parts of a comparable corpus C is then defined as follows: 2 Ptp∈Cpv T ranstf idf (tp, Csv) Dicetf idf (Cp, Cs) = Ptp∈Cpv tf idf(tp) + Pts∈Csv tf idf(ts)
The experiments described in the next section will be performed with the two comparability measures defined here.
Experiments and Results
Taking CorpusPedia as input source, we performed several experiments to build different comparable corpora for three language pairs, namely Portuguese-Spanish,
Portuguese-English, and Spanish-English. These corpora were built using the two strategies described in Section 2 and five domain specific seed terms (in the three languages) considered as representative of five domain topics: “Archaeology”, “Linguistics”, “Physics”, “Biology”, and “Sport”.
Table 1 shows the (binary and tf idf) Dice scores obtained from measuring the comparability degree of 30 different comparable corpora. For each corpus, the table also shows the size (in Mb) of its two parts. In particular, the first column introduces the two languages of the corpus (pt = Portuguese, sp = Spanish, en = English) and the type of strategy (aligned or not aligned) used to build it. In the second and third columns, we show the two Dice scores. The forth column shows the size of the two parts of the corpus, and the last column contains the two seed terms employed to generate the corpus. In Table 2, we show the Dice scores as well as the size of nine pairs of monolingual corpora randomly generated from Wikipedia.
We can observe first that there are significant differences in terms of comparability between the Dice scores in Table 1 and those obtained from the randomly generated monolingual pairs in Table 2. It follows that corpora built by means of our strategies (not aligned and aligned) are actually comparable. Then, we should note that in the comparable corpora of Table 1, the Dice scores based on tf idf are about 70 % higher than those based on the binary function. By contrast, in randomly generated corpora (Table 2), there are no significant differences between Dicebin and Dicetd idf . It means that our tf idf makes the Dice similarity score higher if the two evaluated corpus parts are actually comparable.
As it was expected, not-aligned corpora tend to be larger than the aligned ones. However, if we just compare the smallest parts of each corpus, the differences are not very important: the smallest parts of not-aligned corpora are only 15 % larger than those of aligned corpora. This is in accordance with the fact that aligned corpora are more balanced in terms of size, since no part is much larger than the other one. As far the corpus size is concerned, let us note that, in average, English parts are clearly larger than the Spanish ones, which are slightly larger than the Portuguese ones. In general, English arCorpora pt-sp (not aligned) pt-en (not aligned) sp-en (not aligned) pt-sp (aligned) pt-en (aligned) sp-en (aligned) pt-sp (not aligned) pt-en (not aligned) sp-en (not aligned) pt-sp (aligned) pt-en (aligned) sp-en (aligned) pt-sp (not aligned) pt-en (not aligned) sp-en (not aligned) pt-sp (aligned) pt-en (aligned) sp-en (aligned) pt-sp (not aligned) pt-en (not aligned) sp-en (not aligned) pt-sp (aligned) pt-en (aligned) sp-en (aligned) pt-sp (not aligned) pt-en (not aligned) sp-en (not aligned) pt-sp (aligned) pt-en (aligned) sp-en (aligned) pt-sp (not aligned) pt-en (not aligned) sp-en (not aligned) pt-sp (aligned) pt-en (aligned) sp-en (aligned) .068 .041 .090 .179 .127 .181 .078 .054 .074 .140 .128 .150 .200 .123 .270 .237 .178 .220 .130 .102 .068 .197 .186 .213 .083 .026 .047 .175 .189 .206 .111 .069 .109 .185 .161 .194
(bin)
Dice (tf-idf ) .086 .067 .140 .199 .140 .226 .129 .136 .170 .214 .194 .257 .374 .287 .403 .390 .348 .387 .227 .193 .129 .328 .308 .294 .148 .085 .136 .266 .334 .290 .192 .153 .195 .279 .264 .290
Size (in Mb) 0.6Mb/3.4Mb 0.6Mb/8.4Mb 0.4Mb/8.4Mb 0.4Mb/0.2Mb 0.4Mb/1.1Mb 2.0Mb/2.9Mb 0.8Mb/1.7Mb 0.8Mb/5.1Mb 1.7Mb/5.1Mb 0.6Mb/0.8Mb 0.5Mb/1.2Mb 0.9Mb/1.7Mb 4.4Mb/4.8Mb 4.4Mb/12Mb 4.8Mb/12Mb 3.6Mb/4.7Mb 3.8Mb/11Mb 3.4Mb/7.6Mb 2.4Mb/1.5Mb 2.4Mb/9.4Mb 1.5Mb/9.4Mb 1.6Mb/2.8Mb 1.8Mb/4.5Mb 0.9Mb/1.3Mb 11Mb/35Mb 11Mb/333Mb 35Mb/333Mb 9.7Mb/15Mb 11Mb/20Mb 20Mb/29Mb 3.8Mb/9.3Mb 3.8Mb/73Mb 9.3Mb/73Mb 3.2Mb/4.7Mb 3.5Mb/7.6Mb 6.2Mb/8.5Mb
Arqueologia, Archaeology Arqueolog´ıa, Archaeology Arqueologia, Arqueolog´ıa Arqueologia, Archaeology Arqueolog´ıa, Archaeology
Lingu´ıstica, Linguistics Lingu¨´ıstica, Linguistics Lingu´ıstica, Lingu¨´ıstica Lingu´ıstica, Linguistics Lingu¨´ıstica, Linguistics
F´ısica, Physics F´ısica, Physics F´ısica, F´ısica F´ısica, Physics F´ısica, Physics
Biologia, Biology Biolog´ıa, Biology Biologia, Biolog´ıa Biologia, Biology Biolog´ıa, Biology
Desporto, Sport Deporte, Sport Desporto, Deporte Desporto, Sport Deporte, Sport
Cuadro 1: Dice similarity between several comparable corpora in Portuguese, Spanish, and English.
Corpora pt-sp1 (random) pt-en1 (random) sp-en1 (random) pt-sp2 (random) pt-en2 (random) sp-en2 (random) pt-sp3 (random) pt-en3 (random) sp-en3 (random)
(bin)
Dice (tf-idf ) .012 .003 .003 .016 .017 .017 .008 .001 .005 .012 .003 .003 .014 .014 .015 .006 .001 .005
Size (in Mb) 2.2Mb/0.9Mb 2.2Mb/0.4Mb 0.9Mb/0.4Mb 1.5Mb/3.0Mb 1.5Mb/42Mb 3.0Mb/42Mb 0.2Mb/0.5Mb 0.2Mb/1.4Mb 0.5Mb/1.4Mb Cuadro 2: Dice similarity between randomly generated pairs of monolingual corpora. ticles tend to have more words than Spanish and Portuguese articles. As it was suggested by one of the reviewers of the article, one of the reasons for the difference in size in the case of aligned corpora is that Spanish and Portuguese entries seem to be summaries of the English ones. So, to increase comparability between an aligned pair of articles, the longer article could be shortened by removing those parts which are not present in the other language, obtaining, this way, a more comparable pair of articles.
Finally, as it was expected, aligned corpora are significantly more comparable (i.e., higher Dice coefficient) than not-aligned corpora. In average, Dicetd idf increases 80 % the comparability of aligned-corpora with regard to not-aligned ones. So, considering that aligned corpora only decreases 15 % in size in relation to not-aligned corpora, we can conclude that the aligned strategy seems to be more appropriate to build comparable corpora from Wikipedia.
Conclusions and Future Work
The emergence of multilingual resources, such a Wikipedia, makes it possible to design new methods and strategies to compile corpus from the web, methods that are more efficient and powerful than the traditional ones. In particular, the semi-structured information underlying Wikipedia turns out to be very useful to build comparable corpora. In this article, we proposed two strategies to build comparable corpora from Wikipedia and a way to measure their degree of comparability. The experiments led us to conclude that corpora aligned article by article are more comparable than not aligned corpora.
Besides, they consist of two balanced corpus parts in terms of size. Finally, they are not much smaller than not aligned corpora.
In future work, we will be focused on how
to improve the strategies to build
comparable corpora by extending coverage (more
articles) without losing comparability. For this
purpose, we will test and evaluate
techniques to expand categories using a list of
similar terms identified as hyponyms or
cohyponyms of the source category. In order to
find hyponyms and co-hyponyms of a term, it
will be required to build an ontology of
categories using the semi-structured information
of Wikipedia
Bibliograf´ıa Adafre, S.F. y M. de Rijke. 2006. Finding similar sentences across multiple languages in wikipedia. En 11th Conference of the European Chapter of the Association for
Computational Linguistics, p´aginas 62–69.
Adar, Eytan, Michael Skinner, y Daniel S.
Weld. 2009. Information arbitrage across multi-lingual wikipedia. En Second ACM International Conference on Web Search and Data Mining , WSDM.
Chernov, Sergey, Tereza Iofciu, Wolfgang
Nejdl, y Xuan Zhou. 2006. Extracting semantic relationships between wikipedia categories. En SemWiki2006 - From Wiki to Semantics, Budva, Montenegro. de Melo, Gerard y Gerhard Weikum. 2010.
Menta: inducing multilingual taxonomies from wikipedia. En Proceedings of the 19th ACM international conference on Information and knowledge management,
CIKM ’10, p´aginas 1099–1108.
Filatova, Elena. 2009. Directions for Exploiting Asymmetries in Multilingual Wikipedia. En CLEAWS3, p´aginas 30–37, Colorado.
Gamallo, Pablo y Isaac Gonz´alez. 2010. Wikipedia as a multilingual source of comparable corpora. En LREC 2010 Workshop on Building and Using Comparable Corpora, p´aginas 19–26, Valeta, Malta.
Gamallo, Pablo y Jos´e Ramom Pichel.
2008. Learning Spanish-Galician Translation Equivalents Using a Comparable Corpus and a Bilingual Dictionary. LNCS, 4919:413–423.
Li, Bo y Eric Gaussier. 2010. Improving corpus comparability for bilingual lexicon extraction from comparable corpora. En 20th International Conference on Computational Linguistics (COLING 2010, p´aginas 644–652.