A Siamese network framework for bank intelligent Q&A prediction

Abstract

With the development of financial technologies and the explosive growth of data, the intelligent customer services have attracted a considerable attention of academic and industrial experts. The calculation of question similarity has become a key to the question and answer (Q&A) of financial intelligent customer service. In this work, we propose a Siamese network called W2V-Siamese-BiLSTM for bank Q&A prediction. Based on the Word2vec text information vectorization, the pre-training model of embedding layer is established, and two bidirectional long short-term memory (Bi-LSTM) networks are introduced to encode the upper layer input. The weights are shared in the encoding layer, and finally the Manhattan distance is used in the similarity calculation layer. The experimental results show that the supervised similarity calculation framework proposed in this work has good applicability in real financial Q&A. The accuracy of the proposed method is 81.2%. The analysis further highlights the effectiveness and superiority of supervised learning models in the financial field.