A monthly overview of recent academic research about Wikipedia and other Wikimedia projects, also published as the Wikimedia Research Newsletter.

Reviewed by Tilman Bayer

Flashcards are a popular method for memorizing information. A paper^[1] by six Zurich-based researchers, presented earlier this month at the annual AAAI conference, describes a tool to automatically extract flashcards from Wikipedia articles, aiming "to make independent education more attractive to a broader audience."

A proof-of-concept version is available online, with results available for export in a format that can be used with the popular flashcard software Anki. User can choose from four different variants based on either the entire Wikipedia article or just its introductory section.

The researchers emphasize that "generating meaningful flashcards from an arbitrary piece of text is not a trivial problem" (also concerning the computational effort), and that there is currently no single model that can do this. They separate the task into four stages, each making use of existing NLP techniques:

• summarization, to first extract the most relevant information from Wikipedia (the user can also choose to have this step skipped and instead generate flashcards based on the full text)
• answer identification, where a model extracts answer statements from a given sentence based on context information from the surrounding paragraph
• question generation, where a model constructs a question from the statement generated in the previous step, again taking context information from the surrounding paragraph into account
• To improve quality, these are followed by a final filtering step, where a question-answering model tries to reconstruct the answer based on the paragraph from which the question was extracted, and the generated flashcard is discarded if the reconstructed answer does not overlap enough with the pre-generated answer.

Apart from evaluating the results using quantitative text measures, the researchers also conducted a user study to compare the output of their tool to human-generated flashcards from two topic areas, geography and history, rated by helpfulness, comprehensibility and perceived correctness. The "results show that in the case of geography there is no statistically meaningful difference between human-created and our cards for either of the three aspects. For history, the difference for helpfulness and comprehensibility is statistically significant (p < 0.01), with human cards being marginally better than our cards. Neither category revealed a statistically significant difference in perceived correctness." (However, the sample was rather small, with 50 Mechanical Turk users split into two groups for geography and history.)

A quick test of the tool with the article Wikipedia (introduction only) yielded the following result (text reproduced without changes):

Question: What does Wikipedia use to maintain it's [sic] content?

Question: In 2021, where was Wikipedia ranked?

Question: What language was Wikipedia initially available in?

Question: How many articles are in English version of Wikipedia [sic] as of February 2021?

Question: Who hosts Wikipedia?

Question: Whose vision did Time magazine believe made Wikipedia the best encyclopedia in the world?

Question: What is a systemic bias on Wikipedia?

Question: What did Wikipedia receive praise for in the 2010s?

Question: What two social media sites announced in 2018 that they would help users detect fake news by suggesting links to related Wikipedia articles?

• See the page of the monthly Wikimedia Research Showcase for videos and slides of past presentations.
• @WikiResearch, the Twitter feed associated with this monthly research update, celebrated its ninth anniversary today. Over the past 9 years, we have shared on average 1.9 tweets per day about Wikimedia-related research. The feed is also available in syndicated form on Facebook and Mastodon.

Other recent publications that could not be covered in time for this issue include the items listed below. Contributions, whether reviewing or summarizing newly published research, are always welcome.

Compiled by Tilman Bayer and Miriam Redi

A paper in New Media & Society^[2] argues that

"[...] an 'implicit feudalism' informs the available options for community management on the dominant platforms for online communities. It is a pattern that grants user-administrators absolutist reign over their fiefdoms, with competition among them as the primary mechanism for quality control, typically under rules set by platform companies.

[...] the online encyclopedia Wikipedia operates through a sophisticated democracy among active volunteers. Wikipedia also possesses a widely acknowledged benevolent dictator in the person of founder Jimmy Wales [...] Implicit feudalism has reigned over the dominant platforms for online communities so far, from the early BBSes to AI-enabled Facebook Groups. Peer-production practices surrounding free/open-source software and Wikipedia also exhibit it.

[....] The feudal pattern has by and large been written into the default behaviors of online-community platforms. Exceptions like Wikipedia and Debian have required considerable, intentional effort to counteract the implicit feudalism of their tools’ defaults."

From the abstract:^[3]

"Using a novel technique, a massive database of qualitatively described citations, and machine learning algorithms, we analyzed 1 923 575 Wikipedia articles which cited a total of 824 298 scientific articles in our database and found that most scientific articles cited by Wikipedia articles are uncited or untested by subsequent studies, and the remainder show a wide variability in contradicting or supporting evidence. Additionally, we analyzed 51 804 643 scientific articles from journals indexed in the Web of Science and found that similarly most were uncited or untested by subsequent studies, while the remainder show a wide variability in contradicting or supporting evidence."

From the abstract:^[4]

"Collecting supporting evidence from large corpora of text (e.g., Wikipedia) is of great challenge for open-domain Question Answering (QA). Especially, for multi-hop open-domain QA, scattered evidence pieces are required to be gathered together to support the answer extraction. In this paper, we propose a new retrieval target, hop, to collect the hidden reasoning evidence from Wikipedia for complex question answering. Specifically, the hop in this paper is defined as the combination of a hyperlink and the corresponding outbound link document."

(See also the above review of the "WikiFlash" paper presented at the same conference)

From the abstract:^[5]

"... we employ question answering and entity summarization as extrinsic use cases for a longitudinal study of the progress of KB coverage. Our analysis shows a near-continuous improvement of two popular KBs, DBpedia and Wikidata, over the last 19 years, with little signs of flattening out or leveling off."

See also the video recording of a talk by the authors at Wikidata Workshop 2020.

Presented at the ACM Special Interest Group on Information Retrieval (SIGIR) forum last December, this paper^[6] found that the majority of Question Answering (QA) datasets are based on Wikipedia data.

From the "Evaluation" section of an AAAI'21 paper titled "Identifying Used Methods and Datasets in Scientific Publications":^[7]

"Figure 4c shows the absolute amount of publications for the top four extracted datasets. [...] Another trend is visible for Wikipedia, which has become popular in research on knowledge representation and natural language processing."

The contributions of this paper^[8] include

"a hub of pre-indexed Wikipedia [dumps, of the English and Chinese language versions] at different years with different ranking algorithms as public APIs or cached results". The authors note that "Opendomain QA datasets are collected at different time, making [them depend] on different versions of Wikipedia as the correct knowledge source. [...] Our experiments found that a system’s performance can vary greatly when using the wrong version of Wikipedia. Moreover, indexing the entire Wikipedia with neural methods is expensive, so it is hard for researchers to utilize others’ new rankers in their future research."

This preprint^[9] includes a dataset consisting of 17 conspiracy theory topics from Wikipedia (including e.g. the articles Death of Marilyn Monroe, Men in black, Sandy Hook school shooting) and comes with a content warning ("Note: This paper contains examples of potentially offensive conspiracy theory text").

From the abstract:^[10]

"[We analyze] the Wikipedia edit history to see how spontaneous individual editors are in initiating bursty periods of editing, i.e., spontaneous burstiness, and to what extent individual behaviors are driven by interaction with other editors in those periods, i.e., interaction-driven burstiness. We quantify the degree of initiative (DOI) of an editor of interest in each Wikipedia article by using the statistics of bursty periods containing the editor's edits. The integrated value of the DOI over all relevant timescales reveals which is dominant between spontaneous and interaction-driven burstiness. We empirically find that this value tends to be larger for weaker temporal correlations in the editor's editing behavior and/or stronger editorial correlations. These empirical findings are successfully confirmed by deriving an analytic form of the DOI from a model capturing the essential features of the edit sequence."

(See also our earlier coverage of research on editors' burstiness)