A Scheme for Efficient Question Answering with Low Dimension Reconstructed Embeddings
Happy Buzaaba, Toshiyuki Amagasa · 2021
Question answering (QA) is a fundamental task whose aim is to answer natural language questions. Several embedding based methods that capture semantic similarity between the natural language question and the given answer have been proposed. While these methods achieve good results on the QA task, the high dimensional representations of embeddings comes at a high memory and computational cost. In this work, we propose a scheme where embedding dimensions are reconstructed with a low dimension for solving the question answering task. To be specific, we apply an autoencoder that learns the low dimension properties of the input embedding representations which we then use for measuring similarity between the natural language question and the given answer. We demonstrate through our analysis that with dimensionality reduction, computation time, and memory requirements can be reduced all the while achieving a reasonable performance. Experiments and analysis on insuaranceQA benchmark, show that our proposed method can obtain performance comparable to standard baselines while remaining cost efficient on both time and memory.