Incorporating Statistical Features in Convolutional Neural Networks for Question Answering with Financial Data

The goal of question answering with financial data is selecting sentences as answers from the given documents for a question. The core of the task is computing the similarity score between the question and answer pairs. In this paper, we incorporate statistical features such as the term frequency-inverse document frequency (TF-IDF) and the word overlap in convolutional neural networks to learn optimal vector representations of question-answering pairs. The proposed model does not depend on any external resources and can be easily extended to other domains. Our experiments show that the TF-IDF and the word overlap features can improve the performance of basic neural network models. Also, with our experimental results, we can prove that models based on the margin loss training achieve better performance than the traditional classification models. When the number of candidate answers for each question is 500, our proposed model can achieve 0.622 in Top-1 accuracy (Top-1), 0.654 in mean average precision (MAP), 0.767 in normalized discounted cumulative gain (NDCG), and 0.701 in bilingual evaluation understudy (BLEU). If the number of candidate answers is 30, all the values of the evaluation metrics can reach more than 90%.