|
Abstract:
|
For most businesses, Customer Service Representatives (CSRs) are a common resource deployed via call centers to interact with customers on behalf of their organization. In order to assess the quality of interactions between CSRs and customers, these communications are usually recorded. With a large number of daily audio recordings, mining useful knowledge from the unlimited source of information becomes an increasingly difficult task. In this paper, we extract acoustic features and linguistic features from the source information and propose an aggregated method for voice sentiment recognition. Here, we focus on sentiment labels: negative and nonnegative. For sentiment analysis by linguistic features, we evaluated simple rule-based models (such as VADER), APIs, Support Vector Machines, and RNNs (with and without transfer learning). For the audio sentiment analysis, we did acoustic feature engineering first, then tried various RNNs for sentiment classification. Our results show that the hybrid fusion of linguistic and audio information allows us to exploit complementary information, and combining the two modalities together could enhance sentiment analysis.
|
