Wikipedia is used as a training corpus for many information selection tasks: summarization, question-answering, etc. The information presented in Wikipedia articles as well as the order in which this information is presented, is treated as the gold standard and is used for improving the quality of information selection systems. However, the Wikipedia articles corresponding to the same entry (person, location, event, etc.) written in different languages have substantial differences regarding what information is included in these articles. In this paper we analyze the regularities of information overlap among the articles about the same Wikipedia entry written in different languages: some information facts are covered in the Wikipedia articles in many languages, while others are covered only in a few languages. We introduce a hypothesis that the structure of this information overlap is similar to the information overlap structure (pyramid model) used in summarization evaluation, as well as the information overlap/repetition structure used to identify important information for multidocument summarization. We prove the correctness of our hypothesis by building a summarization system according to the presented information overlap hypothesis. This system summarizes English Wikipedia articles given the articles about the same Wikipedia entries written in other languages. To evaluate the quality of the created summaries, we use Amazon Mechanical Turk as the source of human subjects who can reliably judge the quality of the created text. We also compare the summaries generated according to the information overlap hypothesis against the lead line baseline which is considered to be the most reliable way to generate summaries of Wikipedia articles. The summarization experiment proves the correctness of the introduced multilingual Wikipedia information overlap hypothesis.
CAiRE-COVID: A Question Answering and Query-focused Multi-Document Summarization System for COVID-19 Scholarly Information Management
