Distilling Essence of a Question: A Hierarchical Architecture for Question Quality in Community Question Answering Sites

Community question answering (CQA) sites have grown to be useful platforms where users search for highly specific information to resolve a problem. However, the significant increase in the number of user-generated content with high variance in quality on these sites not only presents challenges for user navigation but also outgrow the community's peer reviewing capacity. This necessitates ways to automatically assess the quality of new questions so as to maintain quality of content served to its users. While existing methods commonly employ social network indicators as features, our model aims to avoid social influence biases arising from these indicators by predicting the quality from semantic evaluation of the question text. Formulation of the proposed model is non-trivial as it requires the extraction of meaningful features from the noisy question text at different granularities while filtering redundant information. In this work, a neural architecture is proposed to address this problem by aggregating the textual features extracted at word- and sentence-level in a hierarchical manner. In addition, a unique attention mechanism that focuses on sentence segments for interpreting a question is developed. This new mechanism employs the global topical information from common problem contexts. The proposed approach is verified on the Stack Overflow question dataset and is shown to outperform existing neural models.