Question Classification using Latent Focus Words Presented by

Question Classification using Latent Focus Words Presented by: Anand Iyer Ritvik Mudur

Motivation

Motivation • Question Answering systems are slowly becoming more prevalent • Web search • Yahoo answers • Forums • Question classification is an integral part of any QA system • Classify each question into a pre-defined set of classes • Fine and coarse • Limits the range of data/corpora to be searched • Makes the problem tractable (in many cases)

Question Focus Words (QFWs) • Word (or phrase) indicating the main focus of the question • Excluding the interrogative words • Example • HUM: ind - Who was the first person to reach the North Pole ? • Focus word: person • Commonly found using heuristics • Templates: • ‘What|Which type|kind|relative of’ • Head word of first NP/VP in the question • Can also “learn” the focus word with a set of labeled “gold” QFWs

Latent QFWs • We could also consider the focus words as “latent” or hidden before training • Use a gradient-descent approach to find the focus words that result in the most confident (and accurate) classifications on our training set

Results • Compared to a baseline system that used only heuristics for QFW selection • Used a perceptron classifier • Begin with a single random initialization • Begin with an ensemble of 20 random initializations • Using a voting system in the first iteration

Conclusion • Comparable results • Many more avenues to explore • • Different training approaches Multiple QFWs Space of latent QFWs Hybrid approach (heuristics + latent focus words) • Exploits ambiguity • Scales easier • To more specific classes (new heuristics can be tedious to introduce) • To more data (do not require “gold” focus words)

Questions

Acknowledgements THANK YOU Mihai Surdeanu for your guidance and support