Question: Which language scientists designed the first competitive quiz-bowl-playing NLP system?
June 19, 2015
Maryland Language Science Center
QANTA, a question-answering system designed by researchers from UMD and the University of Colorado, Boulder, competed against four top-ranked human players in a demonstration quiz bowl match at the 2015 High School National Championship Tournament in Chicago last week. And apparently the match was a nail-biter: QANTA was in the lead going into the final question but the human team rallied to bring the match to a 200-200 tie. The QANTA team includes faculty members Jordan Boyd-Graber (Colorado, formerly of UMD’s iSchool) and Hal Daumé III (CS/UMIACS), and graduate students Alvin Grissom II (Colorado), Mohit Iyyer , Varun Manjunatha, Leonardo Claudino and He He Quiz bowl poses a special NLP challenge when compared to other trivia competitions such as Jeopardy, since contestants can interrupt the question as soon as they think they know the answer. When hearing a question word-by-word, at what point has enough information been received to make an appropriate guess? An incorrect answer based on too little information costs points, but delaying too long allows the other team to answer first. The research behind the system addresses two big challenges: developing a content model that allows the system to make successful predictions, even when information is incomplete; and decision-making processes to identify the best point to answer the question. This research also has valuable applications in other areas - including simultaneous machine translation, another focus of Jordan and Hal’s work which poses closely-related problems. Simultaneous machine translation is difficult when translating between languages with very different word orders: at some point in a sentence you’re forced to either stop and wait for a crucial word, or you have to make a guess and forge ahead, just like in quiz bowl. Congratulations to Jordan, Mohit, Hal and the rest of the team at UMD for an excellent (and fun) demonstration of their research! You can watch Jordan and Mohit explain how QANTA works, or watch the match itself:
