Question-Answer System in the Banking Domain

Question answering systems can take questions in a free form and answer them in a natural language. This task is extremely important for science and business. Successes of many recent papers is primarily associated with natural language understanding. Business is interested in dialogue systems for cutting the operational costs, for example, in customer service. This is due to the fact that dialog systems can respond 24 hours a day and does not make customers wait as people-operated customer service does. Moreover, they do not make grammatical and other mistakes, and can be easily controlled by adding or removing some possible answers. At the moment there is no universal question answering system capable of answering all the possible questions with high accuracy. Usually, each business area is considered separately and the data specifics is taken into account for finding more suitable solutions. In this paper, we consider neural network approaches for solving this problem, in particular using siamese-like ranking models. All experiments are conducted on the corpus of historical data of customers' conversations with bank customer support. The approaches of search for proximity of two questions are explored: questions from the database and customer questions, as well as the answers' ranking based on client's question. For each approach, advantages and disadvantages are revealed. The architecture of the final solution, which was implemented in the bank’ live chat, is described.