Abstract
The expansion of telecommunication services has been met with a rise in the cases of the fraudulent phone calls posing a big threat to individuals and organizations. Conventional techniques of detecting are usually based on offline analysis of full conversations, which restricts their promptness of intervention. In this paper, the author proposes a real-time, turn-taking, fraud detecting system, which is based on pre-trained contextual embeddings in combination with a bi-directional Long Short-Term Memory network in order to model semantic content and temporal dynamics of multi-turn conversations. To detect fraudulent calls, the system progressively changes the probability of a call being a fraud after every conversational turn to allow it to detect a fraud. When tested with a synthetic multi-turn dialogue dataset, it is shown that the proposed BiLSTM using BERT embeddings has a test accuracy of 93.75% and an F 1 score of 93.55, which is higher than the current machine learning and convolutional baselines. The system can note most of the scams during the initial few turns of a call, which offers fast risk evaluation. These findings suggest the usefulness of context-based, progressing modeling to detect fraud in real time and its possibility of practical application.