Real-Time Information Extraction for Phone Review in Car Loan Audit

Abstract

Phone review is important in car loan audits, in which auditors contact applicants to make risk assessments by how applicants act to a sequence of questions. Due to the length of dialogues, auditors tend to miss important details, thus requiring an aiding system to record the dialogues in a compact form. Existing methods that utilize slot-value pairs to track the latest dialogue states fail to record the intermediate process which is critical for risk assessment. In this paper, we propose quadruples which consist of a dialogue act and a triple in a concept graph to represent the dialogue process, and model the dialogue recording task as a quadruple extraction problem for each utterance. To concisely construct quadruples, we convert slot-value pairs into a concept graph by disentangling domains from slots. In order to extract quadruples in real time, we design a model incorporating multi-head cross-attention mechanism and embedding sharing while considering parameter size and inference speed. Experiments on our real-world dialogue dataset show that our model achieves an accuracy of $\sim$82.7% which is similar to the best baseline with only $\sim$30 M parameters while performing real-time inference $\sim$3.6 times faster on an 8-core CPU with $\sim$90 ms per utterance.