A Language Independent Approach to Identify Problematic Conversations in Call Centers

Abstract

Voice based call centers enable customers query for information by speaking to human agents. Most often these call conversations are recorded by call centers with the intent of trying to identify things that can help improve the performance of the call center to serve the customer better. Today the recorded conversations are analyzed by humans by listening to call conversations, which is both time consuming, fatigue prone and not very accurate. Additionally, humans are able to analyze only a small percentage of the total calls because of economics. In this paper which is based on [1], we propose a visual method to identify problem calls quickly. The idea is to sieve through all the calls and identify problem calls, these calls can then be further analyzed by human. We first model call conversations as a directed graph and then identify a directed graph structure associated with a normal call. All call conversations that do not have the structure of a normal call are then classified as being abnormal. We use the speaking rate feature to model call conversation because it can spot potential problem calls. We have experimented on real call center conversations acquired from different call centers and the results are encouraging.