AI has been the talk of the town for quite some time now. AI innovations can revolutionize any industry if rightly applied! Recently, most service organizations have shifted their focus toward adopting AI capabilities to provide quick, efficient, and personalized service journeys. B2C service is not so far away, with an already released AIML accelerator, Incident Classifier, for our customers to classify the incidents, improve routing, and improved SLAs. As a next step towards bringing intelligence and automation to the service center world, B2C Service is proud to unveil our next AIML accelerator, the Live Chat Sentiment Analysis Accelerator for Oracle B2C Service. Here, B2C service customers can proactively help agents, supervisors and end-users to make the service journey pleasant with proper attention at the right time with end-user sentiment trend insights. Result - brand loyalty and stickiness!

The blog gives a sneak peek at the Live Chat Sentiment Analysis Accelerator for Oracle B2C Service, an integration with OCI language Services to leverage its capabilities like custom language models and identify the customer sentiment and manager escalation requests on a live chat.

Note: Please read and understand the open-source licensing terms before configuring the solution.

Please refer to the blog post about Incident Classifier accelerator to learn more about the accelerators.

How does the Live Chat Sentiment Analysis Accelerator work?

The Live Chat Sentiment Analysis Accelerator uses OCI Language Service's custom classification capability to identify the sentiment on each chat response from the end user. Supervisors are presented with a report where they see a list of active chats, insights on the overall sentiment, and the sentiment of each customer response. Additionally, the accelerator can detect when the customer asks for an escalation or wants to talk to a manager as a “Supervisor Ask”. Using the report, the service center supervisor can identify and proactively intervene in the chats where the end user is unhappy before it is too late, evaluate the chats that closed on a positive note, create examples for agents from the positive chats, and appreciate the agents who made the end users while handling multiple chats at the same time. The solution also provides a report on sentiment analysis of wrapped chats for further evaluation and options to provide feedback to the model for finetuning.

Why would I prioritize the Live Chat Sentiment Analysis Accelerator?

Oracle B2C Service customers use the chat feature to help their end users by assisting them in real time. However, the sheer volume of interactions can be overwhelming for the customer support teams.

• When agents handle multiple customers via live chat at the same time, even the best of the agents might miss the small cues from the customers when the customers are getting angry or frustrated. It may be too late by the time the agent reaches out to their supervisors for guidance.
• Periodically, supervisors would like to randomly review the chat conversations to understand the general trend in customer satisfaction over live chats. However, they would not know which live chats require a closer look and the ones that are examples of exemplary work done by the agents.
• When an agent is new to the job, the supervisor would like to keep a close tab on the chat exchanges and must review all the chats, instead of focusing on the ones that require attention and the ones that require an appreciation!

In each of the above instances, it would be helpful for the supervisor to know the customer sentiment as the chat progresses, and proactively intervene to assist the customers promptly, which would also reduce the work stress on agents with timely guidance.

How to make the most out of the Live Chat Sentiment Analysis Accelerator?

The accelerator provides an option where the customers can identify end-user sentiment during a chat conversation and bring any issues to the supervisor's attention. The accelerator is designed specifically for call center use cases with the ability to provide feedback, through which any specific negative sentiment can be captured, which in general context might be positive (Eg: Your competitor has a better service). The solution also,

• Improve end-user satisfaction and retention
• Increase agent efficiency by indicating them with signals and help them to take the next step at the right time while handling multiple chats
• Increase agent efficiency by allowing the supervisor to monitor the agents and provide them with early feedback when something goes wrong in handling a chat.
• Minimize human error by evaluating the negative chats and highlighting negative chats to the supervisor for help.
• Optimize chat auditing efficiency by providing additional support to the supervisors in identifying negative chats.
• Improve resolution time by notifying the supervisor when there is an ask for a supervisor from an end user.

The predicted sentiments can be further utilized to improve the services by providing priority to the customers who were unhappy in the last interaction, or preparing SOPs by taking examples from chats concluded on a positive note, etc. 

Can we have a closer look into the Live Chat Sentiment Analysis Architecture?

The Live Chat Sentiment Analysis accelerator uses the text classification capability of OCI in predicting “Sentiment” and “Supervisor Ask” for any chat response posted by an end user. 

• The accelerator evaluates the chat text from the end user by using an Extension code, which will call the OCI language service via External Objects to get the emotions and supervisor ask.
• Async Custom Process Module (CPM) is used to process the evaluation of prediction results of each chat text and then flags the chat when a “supervisor ask” or “negative emotion” pattern is identified.
• Accelerator creates an initial custom OCI language model using historical end-user chat responses and the corresponding emotions tagged by the EmoRoBERTa model.
• The solution also outlines the mechanism of fine-tuning the initial model in case of incorrect predictions using agent feedback as a way forward.

The following architecture diagram depicts the extension and CPM flow:

How to unlock the potential of Live Chat Sentiment Analysis Accelerator?

Accelerators are the sample code, and integration stack configuration released along with documentation to implement for the customers to solve their problems by integrating to other services using Oracle B2C Service's capabilities under UPL License. (For more details on UPL, see- https://oss.oracle.com/licenses/upl/). The sample code can be configured as it is or customized as per the customer’s unique business use cases and solve the issues specific to them.

Sample code along with configuration guide is available on:

1. Browser UI Admin Page: Users with administrator privilege can access the list of accelerators through the Administrator page. Users can open the link provided to view documentation or move to the download sample code page  
2. Oracle B2C Service Accelerators OTN Site: Users can also access the details through the public-facing Oracle Technical Network site.  

Where can I read more about the topic?

Blog about Incident Classifier Accelerator

OCI Custom Language Classification model documentation

EmoRoBERta Model

B2C Service Accelerators