Predicting Triple Scoring with Crowdsourcing-specific Features - The fiddlehead Triple Scorer at WSDM Cup 2017
This work addresses the problem of improving prediction accuracy for crowdsourced data in information retrieval, but it is incremental as it builds on existing methods with specific feature additions.
The paper tackled predicting crowdsourced relevance scores between persons and professions/nationalities by incorporating features related to task difficulty, which improved the average score difference of the prediction, achieving 4th place in that metric in the WSDM Cup 2017 competition.
The Triple Scoring Task at the WSDM Cup 2017 involves the prediction of the relevance scores between persons and professions/nationalities. The ground truth of the relevance scores was obtained by counting the vote of seven crowdworkers. I confirmed that features related to task difficulty correlate with the discrepancy among crowdworkers' judgement. This means such features are useful for predicting whether a score is in the middle or not. Hence, the features were incorporated into the prediction model of the crowdsourced relevance scores. The introduced features improve the average score difference of the prediction. The final ranking of my prediction was 4th for average score difference and 12th for both accuracy and Kendall's tau.