Humans are very good at identifying noun phrases and their corresponding anaphors (phrases that refer back to another word in the text to avoid repetition such as third person pronouns), but technology cannot accurately and efficiently perform this task. That's where TileAttack comes in. It’s a fun, fast-paced, competitive language game about identifying noun phrases and anaphors that also serves the purpose of strengthening noun phrase detection in machine learning technology. Tile Attack is a part of the DALI project, which was founded in order to analyze disagreements in language interpretation and ambiguity. This project also contains other games such as Phrase Detectives, Lingotorium, and WordClicker Bakery, which are featured on LingoBoingo. By using gamified platforms, researchers in the DALI project can collect a large amount of data and study human and technology identified disagreements.
Tile Attack logo
Screenshot of the DALI Project's website
TileAttack is a good example of a Game with a Purpose (GWAP), a term coined by the founder of Duolingo, Louis von Ahn. GWAPs are games played by humans that produce useful data for research in many different fields such as computer science and linguistics. GWAPS are especially beneficial when researchers need to collect a large amount of data but lack the necessary resources (staff, time, funding) to produce the data on their own.
A YouTube video entitled "TileAttack: A GWAP for gathering mentions" posted by the GamesAndNLP channel discusses the motivations behind creating TileAttack. Researchers wanted to collect data to better understand how humans comprehend and categorize noun phrases, but weren't necessarily able to support the cost of compensation, so they created a game that was fun for participants and would, as a result of playing, produce data in the form of annotations. The data are used to create corpora to train language and machine learning technologies how to understand and deal with noun phrases in language. By understanding how humans use this concept in language, these technologies can perform tasks like sentiment analysis, machine translation, and information extraction.
Participants are shown words and their task is to find noun phrases or mentions that correspond to those words. The following example is given in the tutorial. To complete this section, participants would locate the words or phrases that refer to "music". Both "The music" and "it" refer to "music" in this example, so participants would click those boxes.
When participants finish the tutorial, they progress to the game and are tasked with identifying all noun phrases that occur in given sentences. Participants click on groups of word tokens that make up noun phrases and then press the "Annotate" button on the bottom left of the page to submit. They compete against a second player, and points are given based on how well their answers match up. After participants have annotated all the noun phrases, they press the “Finish” button, view how their answers compare with the second user and see which noun phrases they didn't identify.
After each round is over, participants can see high scores of the month and their ranking in comparison with other participants.
Poesio, M., Chamberlain, J., Bartle, R., Madge, C., Kruschwitz, U., & Paun, S. (n.d.). Games With a Purpose for Corpus Annotation in the DALI Project.