A Knowledge Based Two-Stage Cascade Model for Test-Question/Answer Retrieval

Abstract

Test-question/answer retrieval task has raised higher requirements in terms of accuracy, coverage and semantic understanding. We design a cascade model with two-stage training processes: The first stage uses 41,532 user test-question click records and 207,660 unclick records, which are collected from a designed test-question-answer experimental platform, to generate 200,000 pairwise training dataset to train a deep learning model, which could improve generalization ability. The second stage combines the output of the first stage with structural knowledge as new features to train a logistic regression for selecting the results from the candidates with higher accuracy, the training dataset is generated by manually annotating 20,000 test-question samples. The structural knowledge is also manually extracted from the samples for generating a small knowledge graph, and on this condition, we design knowledge features. Experimental results show that the proposed model outperforms the state-of-the-art algorithms, among which the cascading model contributes 3% improvement and the knowledge features contribute 1% improvement.