Developing a Multilingual Corpus of Wikipedia Biographies

Proceedings of Recent Advances in Natural Language Processing, pages 285–294 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_032 285 Developing a Multilingual Corpus of Wikipedia Biographies Hannah Devinney12, Anton Eklund1, Igor Ryazanov1, Jingwen Cai1 1 Umeå University, Department of Computing Science 2 Umeå Centre for Gender Studies Umeå, Sweden [hannahd, antone, igorr, jingwenc]@cs.umu.se Abstract For many languages, Wikipedia is the most accessible source of biographical information. Studying how Wikipedia describes the lives of people can provide insights into societal biases, as well as cultural differences more generally. We present a method for extracting datasets of Wikipedia biographies. The accompanying codebase is adapted to English, Swedish, Rus- sian, Chinese, and Farsi, and is extendable to other languages. We present an exploratory analysis of biograph- ical topics and gendered patterns in four lan- guages using topic modelling and embedding clustering. We find similarities across lan- guages in the types of categories present, with the distribution of biographies concentrated in the language’s core regions. Masculine terms are over-represented and spread out over a wide variety of topics. Feminine terms are less fre- quent and linked to more constrained topics. Non-binary terms are nearly non-represented. 1 Introduction Wikipedia’s decentralised organisation and inde- pendent communities in different languages have led it to be considered a ‘global repository of knowl- edge’ (Callahan and Herring, 2011). Easily and openly accessible, in many language environments it has displaced more traditional and specialised re- sources as a ‘default’ encyclopedic source, shaping the language landscape. This effect is noticeable in NLP research and development, where Wikipedia is a staple training data source for language mod- els that imitate Wikipedia when generating texts or making writing suggestions. When it comes to biographical information, Wikipedia has become a primary source of refer- ence, especially outside of education systems. Be- cause of the near-monopoly that Wikipedia has on public knowledge in these environments, it can de- fine which persons are perceived as notable, which aspects of their lives deserve a mention, and how the persons are presented. The community guide- lines of Wikipedia are built around the concept of ‘neutrality’, but the content is still inevitably shaped by societal biases, such as gender gaps (Hube, 2017). Besides the content of the biographi- cal articles, Wikipedia also shapes the expectations the reader has in terms of format, language, style and inclusion. A ‘biography’ can come to invoke a Wikipedia-like structure, becoming a commonly accepted way of summarising a person’s life. Most automatically extracted datasets consists of Wikipedia backup dumps or rely heavily on the connection to Wikidata. Using Wikidata, however, makes it harder to modify or update the dataset, or replicate it for another domain. Manually col- lected datasets, on the other hand, are limited in size. They are almost inevitably biased towards longer, popular or better-categorised articles be- cause poor categorisation prevents other articles to be discovered in the first place. These problems become even more apparent when creating a mul- tilingual dataset. While different editions share a general article structure and templates, they are far …

Content truncated to 500 words