Language-Independent Methods for Identifying Cross-Lingual Similarity in Wikipedia

The diversity and richness of multilingual information available in Wikipedia have increased its significance as a language resource. The information extracted from Wikipedia has been utilised for many tasks, such as Statistical Machine Translation (SMT) and supporting multilingual information access. These tasks often rely on gathering data from articles that describe the same topic in different languages with the assumption that the contents are equivalent to each other. However, studies have shown that this might not be the case. Given the scale and use of Wikipedia, there is a need to develop an approach to measure cross-lingual similarity across Wikipedia. Many existing similarity measures, however, require the availability of "language-dependent" resources, such as dictionaries or Machine Translation (MT) systems, to translate documents into the same language prior to comparison. This presents some challenges for some language pairs, particularly those involving "under-resourced" languages where the required linguistic resources are not widely available. This study aims to present a solution to this problem by first, investigating cross-lingual similarity in Wikipedia, and secondly, developing "language-independent" approaches to measure cross-lingual similarity in Wikipedia. Two main contributions were provided in this work to identify cross-lingual similarity in Wikipedia. The first key contribution of this work is the development of a Wikipedia similarity corpus to understand the similarity characteristics of Wikipedia articles and to evaluate and compare various approaches for measuring cross-lingual similarity. The author elicited manual judgments from people with the appropriate language skills to assess similarities between a set of 800 pairs of interlanguage-linked articles. This corpus contains Wikipedia articles for eight language pairs (all pairs involving English and including well-resourced and under-resourced languages) of varying degrees of similarity. The second contribution of this work is the development of language-independent approaches to measure cross-lingual similarity in Wikipedia. The author investigated the utility of a number of "lightweight" language-independent features in four different experiments. The first experiment investigated the use of Wikipedia links to identify and align similar sentences, prior to aggregating the scores of the aligned sentences to represent the similarity of the document pair. The second experiment investigated the usefulness of content similarity features (such as char-n-gram overlap, links overlap, word overlap and word length ratio). The third experiment focused on analysing the use of structure similarity features (such as the ratio of section length, and similarity between the section headings). And finally, the fourth experiment investigates a combination of these features in a classification and a regression approach. Most of these features are language-independent whilst others utilised freely available resources (Wikipedia and Wiktionary) to assist in identifying overlapping information across languages. The approaches proposed are lightweight and can be applied to any languages written in Latin script; non-Latin script languages need to be transliterated prior to using these approaches. The performances of these approaches were evaluated against the human judgments in the similarity corpus. Overall, the proposed language-independent approaches achieved promising results. The best performance is achieved with the combination of all features in a classification and a regression approach. The results show that the Random Forest classifier was able to classify 81.38% document pairs correctly (F1 score=0.79) in a binary classification problem, 50.88% document pairs correctly (F1 score=0.71) in a 5-class classification problem, and RMSE of 0.73 in a regression approach. These results are significantly higher compared to a classifier utilising machine translation and cosine similarity of the tf-idf scores. These findings showed that language-independent approaches can be used to measure cross-lingual similarity between Wikipedia articles. Future work is needed to evaluate these approaches in more languages and to incorporate more features.