Our project has so far transformed data from multiple online enthusiast communities as well as other data-centric projects into the RDF data format and it is now available online as Linked Open Data. The next step is matching the corresponding entities in each source and merge the properties that each source provides.Continue reading
Data quality and ground truth
After taking a six month break due to an internship, I restarted my work as a student assistant for the Japanese Visual Media Graph project in April 2020.
Currently, my main occupation is in the field of data quality control. After getting lots of data from different fan communities, the quality of said data needs to be checked against other sources to make sure there aren’t any errors adopted into the project’s database. To get started with this task, it was decided to first check several small data samples from different providers to enable an easier determination of the duration, effort, and expectable problems and results that a wider data quality control would entail.
I received the first two data samples from two different fan sites, both containing twenty entries of anime with several properties for me to check; those properties were, for example, the Japanese and English titles of the anime, the producing studio’s name in Japanese and English, the release date of the first episode, and the overall episode count or the completeness of a series. The properties in the samples depended on the usage of properties by the fan communities the data came from, and my task was to check if the entries were correct or if they contained some errors. To prove something, I had to find a source of ground truth for it, which would occasionally prove to be some kind of a challenge. Of course, the anime on DVD would actually be the best source of ground truth, but since the resources for this simply didn’t exist, I relied on other sources. A valid source of ground truth would, for example, be the opening or ending sequence of an anime, preferably found on YouTube or a legal streaming source like Netflix or Crunchyroll. An image of the DVD-case of an anime would also be a usable source for ground truth.
I worked with the data in an excel sheet, marking the correctness of the respective properties accordingly and adding screenshots and links to my sources of ground truth.
I encountered a noticeable difference in finding ground truth for the different properties. Finding proof for the Japanese or English anime titles almost never posed a problem; they usually could be found in the opening sequences or on DVD-cases. The year of first release could also be usually seen in the opening or ending sequences. The exact date was however sometimes quite difficult to proof. While I tried to use the Japanese Amazon Prime at first, it proved to be not reliable enough. Most of the time I could only return to the Media Arts Database to find proof for an exact date of release.
The names of the producing studios could usually be found easily inside the opening and ending sequence of the respective anime; however I sometimes encountered the problem that the studio didn’t write its own name in katakana, like provided by the fan-based data. While I usually could validate that it was indeed the correct studio, the spelling couldn’t be found in any official source for ground truth. I always marked those occurrences as “correct but strange” and left it open for further decisions.
A complicated property was the completeness of an anime. Whether or not this is a usable or provable property remains to be discussed.
After having checked those first two data samples, I can state that there was a lot of correct data, but also errors of different types. How to deal with them is also currently a point of discussion. The next samples from new sources will surely bring even more new experiences and insights.