ABSTRACT

Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 9.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255

Building machines that are capable of answering natural language questions is one of the biggest challenges in natural language processing and has a history dating back to the late 1950s [634]. A renewed interest in question-answering started in 1999 with a series of NIST-sponsored TREC Question-Answering (QA) Tracks, large-scale evaluations of domainindependent question-answering1 that produced a few high performance question-answering systems capable of answering factoid and list questions

[525, 478, 283]. In 2011, IBM’s open-domain question-answering system, Watson [215], beat the two highest ranked Jeopardy!2 players, marking a significant milestone in the more than 60-year quest to create a viable questionanswering machine.