Online Call Scene Segmentation of Contact Center Dialogues based on Role Aware Hierarchical LSTM-RNNs
Ryo Masumura, Setsuo Yamada, Tomohiro Tanaka, Atsushi Ando, Hosana Kamiyama, Yushi Aono · 2018
This paper proposes novel online call scene segmentation methods for contact center dialogues between an operator and a customer. Call scene segmentation is useful for many useful tasks such as summarization and information retrieval. In addition, its online implementation is beneficial in constructing operator assist systems. We define call scene segmentation as utterance-level sequential labeling with five tag types: opening, requirement confirmation, response, customer confirmation, and closing. In order to perform online call scene segmentation precisely, it is essential to capture interactions between the operator and the customer. In fact, dialogues usually include several sets of interactions; for example, the customer raises a question and the operator answers it, or the customer reveals own personal information and the operator repeats it. In addition, each scene shows some tendencies of interactions; for example, customer confirmation includes operator repetitions more frequently than other scenes. To capture the interactions, we present novel fully neural network based methods called role-aware hierarchical long short-term memory recurrent neural networks. The proposed methods can capture long-range interactive information by simultaneously dealing with, in an online manner, both the sequence of sentences and the sequence of speaker role labels. An experiment on Japanese simulated contact center dialogue data sets demonstrates the effectiveness of the proposed methods.