COnVoy: A Contact Center Operated Pipeline for Voice of Customer Discovery
Authors: Rishabh Tripathi = Digvijay Anil Ingle (equal contribution) , Ayush Kumar, Cijo George, Jithendra Vepa
This blog is intended to abridge one of the research works by Observe.AI accepted at Interspeech 2023, Dublin, Ireland
The increasing volume of customer service interactions in contact centers worldwide has the potential to act as a valuable source of information to understand customer preferences and pain-points. This blog discusses an unsupervised pipeline, COnVoy (see Figure-1), which utilizes this data to systematically track the reasons behind the conversations and enable Voice of Customer (VoC) Discovery. The insights gained from this pipeline offer a bird’s-eye view of the business through their contact center conversations, allowing business leaders to devise strategic measures for providing high-quality and faster resolutions to customers, leading to improved customer satisfaction.
One of the most important aspects of effective contact center operations is the ability to identify and track the reasons for customer service interactions, typically characterized by:
- Diversity:
Customers may reach out to contact centers for various reasons such as product or service enquiries, technical support, feedback or complaints, etc. Contact centers tend to watch out for the top-reasons that their customers are reaching out for. - Dynamicity:
Most common issues that customer’s reach out for in this week may not be the same as that in subsequent weeks. In other words, there could be perturbations in the trend of distribution of top-reasons of customer interactions across time.
With the proposed pipeline, the following objectives are achieved:
- Detecting the reason for an interaction and consequently generating an executive summary to provide a bird’s eye view of the business over a given period of time.
- Surfacing newly emerging and out-of-distribution reasons for which customers are calling.
- Ability to map the reason of the call to evidence in the conversation for explainability and to enable further investigation into conversation.
Architecture
- Reason-utterance detector block consists of a text classifier
aimed at predicting whether a given turn in the text transcript
represents the reason for the call or not. For this, we fine-tune DistilRoBERTa-base on a proprietary conversational dataset. Accuracy of the classifier on the test set is 88.23%. - Rewriting module involves generating a synthesized version of the reason for the call, using an in-house LLM (Large Language Model). This step is essential since the output of the reason-utterance detected in the previous step is an excerpt from the actual spoken conversation which comprises of several nuances such as ASR errors, disfluencies, co-references, etc.
- Topic Modeling is performed over the synthesized call-reasons
using BERTopic to cluster them into coherent groups. We
observe a coherence score of 0.53 on a proprietary test set. - Describing the coherent groups involves generating a short description (using an in-house LLM) for the coherent groups identified in the previous step. This short description helps business stakeholders in getting a quick understanding of the data.
Discovering Voice of Customer (VoC)
COnVoy can help in discovering Voice of Customer (VoC) insights in contact-centers by generating an Insights Report that provides a distribution of calls across different call-reason themes, and a week on week analysis of distribution of call reasons to identify newly emerging customer issues
For example, in Figure-3, we observe that the occurrence of the theme Login Issues has almost doubled week on week. This might be insightful to contact centers as they might want to reach out to their technology teams to investigate for the cause of this and decide on an appropriate course of action. Furthermore, we also observe a newly emerging theme Unsubscribe in the current week that was not present in the previous week. Implying that a significant number of customers are requesting to get unsubcribed from their service which might be a critical information for their business. Hence, they would want to further deep dive into its root cause.
Additionally, COnVoy also provides a capability to customers to further drill down into the evidences in call transcripts for explainability and to better understand context around these issues. Such insights report help contact centers make data-backed decisions and take appropriate actions based on customer feedback.
TLDR; COnVoy enables Voice of Customer Discovery from contact center conversations by surfacing key-insights around specific reasons for which their customers are reaching out. It helps identify newly emerging themes of customer issues, thus helping contact centers in reducing the time-to-action to employ strategic measures to fix them. Furthermore, it can map the identified call-reasons to specific evidences in the conversation thus providing explainability and empowering contact centers to perform deeper analysis wherever necessary. In short, the insights surfaced by our pipeline can be used by contact centers to take strategic measures to improve quality of service and operational efficiency, and thus help drive better customer satisfaction.
Thanks for taking interest in patiently reading this blog! :)