Yasuaki Yoshida, Hideyuki Maeda, Tatsuhiro Niwa, Sumio Fujita

The 12th NTCIR Conference (NTCIR-12)

June 07, 2016

Yahoo Japan Search Technology (YJST) team participated in the Query Understanding subtask of NTCIR-12 IMine-2. We explored various search log mining techniques to discover subtopics against the given original topics. For Vertical Identification, we trained a Gradient Boosted Decision Tree (GBDT) learner to identify a vertical label to each subtopic using several complex features including topical probabilities based on random walks on click graphs, and query distribution analyses through several commercial vertical search services and so on. Our best official submission run of subtopic mining achieves higher D#-nDCG@10 score than the average, but below the median of the best runs of all the participating team. In other measures such as QU-score or V-score, our best result performed as poorly as below the median. Although our task campaign was clearly not successful enough to confirm the adequacy of our service solutions, we try to analyze the results and the data as fail- ure analyses are the only way to make a progress towards the future.

Paper : YJST at the NTCIR-12 IMine-2 Task (external link)