View - ResearchGate

82 Socially Intelligent Agentsemotion (non-angry, non-happy, etc.) was 85–92%. The important question ishow to combine opinions of the experts to obtain the class of a given sample.A simple and natural rule is to choose the class with the expert value closest to1. This rule gives a total accuracy of about 60% for the 10-neuron architecture,and about 53% for the 20-neuron architecture. Another approach to rule selectionis to use the outputs of expert recognizers as input vectors for a new neuralnetwork. In this case, we give the neural network the opportunity to learn itselfthe most appropriate rule. The total accuracy we obtained 9 was about 63%for both 10- and 20-node architectures. The average accuracy for sadness wasrather high (∼76%). Unfortunately, the accuracy of expert recognizers was nothigh enough to increase the overall accuracy of recognition.4. DevelopmentThe following pieces of software were developed during the second stage:ERG – Emotion Recognition Game; ER – Emotion Recognition Software forcall centers; and SpeakSoftly – a dialog emotion recognition program. Thefirst program was mostly developed to demonstrate the results of the above research.The second software system is a full-fledged prototype of an industrialsolution for computerized call centers. The third program just adds a differentuser interface to the core of the ER system. It was developed to demonstratereal-time emotion recognition. Due to space constraints, only the second softwarewill be described here.4.1 ER: Emotion Recognition Software For Call CentersGoal. Our goal was to create an emotion recognition agent that can processtelephone quality voice messages (8 kHz/8 bit) and can be used as a part of adecision support system for prioritizing voice messages and assigning a properagent to respond the message.Recognizer. It was not a surprise that anger was identified as the most importantemotion for call centers. Taking into account the importance of angerand the scarcity of data for some other emotions, we decided to create a recognizerthat can distinguish between two states: “agitation” which includesanger, happiness and fear, and “calm” which includes normal state and sadness.To create the recognizer we used a corpus of 56 telephone messagesof varying length (from 15 to 90 seconds) expressing mostly normal and angryemotions that were recorded by eighteen non-professional actors. Theseutterances were automatically split into 1–3 second chunks, which were thenevaluated and labeled by people. They were used for creating recognizers 10using the methodology developed in the first study.