Cross-lingual semantic relatedness using encyclopedic knowledge

Abstract

In this paper, we address the task of crosslingual semantic relatedness. We introduce a method that relies on the information extracted from Wikipedia, by exploiting the interlanguage links available between Wikipedia versions in multiple languages. Through experiments performed on several language pairs, we show that the method performs well, with a performance comparable to monolingual measures of relatedness. MotivationGiven the accelerated growth of the number of multilingual documents on the Web and elsewhere, the need for effective multilingual and cross-lingual text processing techniques is becoming increasingly important. In this paper, we address the task of cross-lingual semantic relatedness, and introduce a method that relies on Wikipedia in order to calculate the relatedness of words across languages. For instance, given the word factory in English and the word lavoratore in Italian (En. worker), the method can measure the relatedness of these two words despite the fact that they belong to two different languages.Measures of cross-language relatedness are useful for a large number of applications, including cross-language information retrieval (Nie et al., 1999;Monz and Dorr, 2005), cross-language text classification (Gliozzo and Strapparava, 2006), lexical choice in machine translation (Och and Ney, 2000;Bangalore et al., 2007), induction of translation lexicons (Schafer and Yarowsky, 2002), cross-language annotation and resource projections to a second language (Riloff et al., 2002;Hwa et al., 2002;Mohammad et al., 2007).The method we propose is based on a measure of closeness between concept vectors automatically built from Wikipedia, which are mapped via the Wikipedia interlanguage links. Unlike previous methods for cross-language mapping, which are typically limited by the availability of bilingual dictionaries or parallel texts, the method proposed in this paper can be used to measure the relatedness of word pairs in any of the 250 languages for which a Wikipedia version exists.The paper is organized as follows. We first provide a brief overview of Wikipedia, followed by a description of the method to build concept vectors based on this encyclopedic resource. We then show how these concept vectors can be mapped across languages for a cross-lingual measure of word relatedness. Through evaluations run on six language pairs, connecting English, Spanish, Arabic and Romanian, we show that the method is effective at capturing the cross-lingual relatedness of words, with results comparable to the monolingual measures of relatedness.

Comparing Semantic Relatedness between Word Pairs in Portuguese Using Wikipedia

“…Our correlation scores are close to the scores achieved by Hassan and Milhalcea [7] when using the MC30 dataset to evaluate a method based on Explicit Semantic Analysis (ESA). In that work, the authors achieved a Pearson correlation of r = .58 and a Spearman correlation of ρ = .75 for the English Wikipedia.…”

Comparing Semantic Relatedness between Word Pairs in Portuguese Using Wikipedia

“…Our correlation scores are close to the scores achieved by Hassan and Milhalcea [7] when using the MC30 dataset to evaluate a method based on Explicit Semantic Analysis (ESA). In that work, the authors achieved a Pearson correlation of r = .58 and a Spearman correlation of ρ = .75 for the English Wikipedia.…”

Combining Heterogeneous Knowledge Resources for Improved Distributional Semantic Models

“…This different behavior of the proposed measure can be attributed to the fact that less weight is given to overlapping low-weight vector components (compared to the other vector measures used here). Our results with this configuration are comparable to (with an advantage for languages with smaller Wikipedias) the Spearman correlation values reported using concept vector based SR with Wikipedia for English (0.75) [9]; for German (0.65) [34]; and for Arabic (0.26), Romanian (0.28) and Spanish (0.50) [13]. However, differences in the Wikipedia versions, preprocessing, etc.…”

Towards Semantic Language Classification: Inducing and Clustering Semantic Association Networks from Europarl

“…Doing so for each language L k gives 20 different correlation coefficients, one for each network version of English, shown in Table 2 We first note that the correlation coefficients differ between network versions of English, where the Italian version exhibits the highest correlation with the (English) human reference, and the Lithuanian version the lowest. Note that Hassan and Mihalcea (2009) obtain a correlation coefficient of 0.55 on the whole WordSimilarity-353 dataset, which is considerably higher than our best score of 0.34. However, first note that our networks, which consist of 5,021 lexical units, are quite small compared to the data sizes that other studies rely on, which makes a comparison highly unfair.…”

Comparing Semantic Relatedness between Word Pairs in Portuguese Using Wikipedia

“…Our correlation scores are close to the scores achieved by Hassan and Milhalcea [7] when using the MC30 dataset to evaluate a method based on Explicit Semantic Analysis (ESA). In that work, the authors achieved a Pearson correlation of r = .58 and a Spearman correlation of ρ = .75 for the English Wikipedia.…”

Combining Heterogeneous Knowledge Resources for Improved Distributional Semantic Models

“…This different behavior of the proposed measure can be attributed to the fact that less weight is given to overlapping low-weight vector components (compared to the other vector measures used here). Our results with this configuration are comparable to (with an advantage for languages with smaller Wikipedias) the Spearman correlation values reported using concept vector based SR with Wikipedia for English (0.75) [9]; for German (0.65) [34]; and for Arabic (0.26), Romanian (0.28) and Spanish (0.50) [13]. However, differences in the Wikipedia versions, preprocessing, etc.…”

Towards Semantic Language Classification: Inducing and Clustering Semantic Association Networks from Europarl

“…Doing so for each language L k gives 20 different correlation coefficients, one for each network version of English, shown in Table 2 We first note that the correlation coefficients differ between network versions of English, where the Italian version exhibits the highest correlation with the (English) human reference, and the Lithuanian version the lowest. Note that Hassan and Mihalcea (2009) obtain a correlation coefficient of 0.55 on the whole WordSimilarity-353 dataset, which is considerably higher than our best score of 0.34. However, first note that our networks, which consist of 5,021 lexical units, are quite small compared to the data sizes that other studies rely on, which makes a comparison highly unfair.…”

Comparing Semantic Relatedness between Word Pairs in Portuguese Using Wikipedia

“…Our correlation scores are close to the scores achieved by Hassan and Milhalcea [7] when using the MC30 dataset to evaluate a method based on Explicit Semantic Analysis (ESA). In that work, the authors achieved a Pearson correlation of r = .58 and a Spearman correlation of ρ = .75 for the English Wikipedia.…”

Combining Heterogeneous Knowledge Resources for Improved Distributional Semantic Models

“…This different behavior of the proposed measure can be attributed to the fact that less weight is given to overlapping low-weight vector components (compared to the other vector measures used here). Our results with this configuration are comparable to (with an advantage for languages with smaller Wikipedias) the Spearman correlation values reported using concept vector based SR with Wikipedia for English (0.75) [9]; for German (0.65) [34]; and for Arabic (0.26), Romanian (0.28) and Spanish (0.50) [13]. However, differences in the Wikipedia versions, preprocessing, etc.…”

Towards Semantic Language Classification: Inducing and Clustering Semantic Association Networks from Europarl

“…Doing so for each language L k gives 20 different correlation coefficients, one for each network version of English, shown in Table 2 We first note that the correlation coefficients differ between network versions of English, where the Italian version exhibits the highest correlation with the (English) human reference, and the Lithuanian version the lowest. Note that Hassan and Mihalcea (2009) obtain a correlation coefficient of 0.55 on the whole WordSimilarity-353 dataset, which is considerably higher than our best score of 0.34. However, first note that our networks, which consist of 5,021 lexical units, are quite small compared to the data sizes that other studies rely on, which makes a comparison highly unfair.…”