Tailored Real-Time Call Summarization System for Contact-Centers

This blog is a flavorful synopsis based on our paper accepted at Interspeech, 2023 — “Tailored Real-Time Call Summarization System for Contact Centers.”

Imagine a bustling contact center, the lifeblood of customer service for a myriad of businesses. Every one of those client calls needs to be summed up neatly — for compliance, as a hot potato pass to the next agent, or simply as a handy reference for future encounters. The result? Contact-center agents spend a considerable amount of their time crafting these critical, yet time-consuming, call summaries. This shores up costs and carves out chunks of their productive time.

Enter the landscape of pre-trained Large Language Models (LLM) that boast call summarization capabilities. But hang on! They gloss over the nitty-gritty of domain-specific information pivotal for contact center businesses and are difficult to customize. In addition, generating summarization after the call still requires the agents to retain the complete context of the call in order to validate the auto-generated summary leading to higher chances of important information being missed or the presence of incorrect information.

In this work, we’ve come up with a hybrid solution that boosts the note-generation capabilities of contact-center-specific LLM and customizes it for the domain-specific needs of contact center call summarization. Let’s dive into the specifics!

The Magic Behind the System

Proposed Workflow: Tailored Real-Time Call Summarization

Sample Instruction for Training Summarization Model. The model is trained to generate text in green with text in red as input.

In the dynamic landscape of Natural Language Processing (NLP), recent advancements have ushered in a wave of innovation aimed at building models capable of understanding and following natural language instructions. Leveraging this progress, our research focuses on fine-tuning an open-source Large Language Model (LLM) using a curated instruction dataset tailored for contact-center call summary generation. Our approach encompasses diverse conversation scenarios across multiple industry verticals, ensuring adaptability to the unique needs of various business sectors. In this blog post, we explore our methodology and the development of an efficient summarization system, enabling real-time streaming predictions while considering specific business intents. The overall methodology is as follows:

1. Dataset Collection: We assembled a curated instruction dataset, meticulously designed for contact center call summary generation, encompassing diverse conversation scenarios from various industries.
2. Fine-Tuning the LLM: We chose an LLM and aligned it with our objective using the paradigm of Supervised Fine-Tuning.
3. Manual Data Annotation: To evaluate the generated summaries accurately, we employed manual data annotation as part of the evaluation process.
4. Real-Time Streaming Optimization: By deliberately using smaller context windows during training, we optimized the model for real-time streaming predictions, improving responsiveness.
5. Custom Business Intent Solution: Instead of retraining the model for individual contact-centers, we allowed users to specify their business intents. We developed an in-memory intent-matching algorithm, capable of identifying specific business intents in real time.
6. Contextual Summarization: Leveraging the in-memory intent-matching algorithm, we dynamically selected context windows for generating contextually fluent summaries.
7. Agent Console Integration: The generated summaries are sent directly to contact center agents’ consoles, providing them with real-time, tailored information to enhance their performance.

Example of intents and their definitions.

A Glimpse at the Results

Sample of Auto-Generated Summary for a 10 min voice call.

Let us dive into the transformative power of auto-generated voice call summaries.

• Effective Generated Summary: Above image presents an example of a generated summary for a 10-minute voice call, effectively capturing essential details such as the call’s purpose, agent greetings, and specific named entities.
• Compliance and Privacy Considerations: The summary intentionally omits any PCI/PII-related information, such as phone numbers, due to compliance requirements. This is achieved by biasing the model during training to avoid producing such sensitive entities.

And its impact on businesses:

• Reduced After-Call Work and Average Call Handling Time: The generated summary offers significant advantages in two primary scenarios. Firstly, it reduces after-call work, an essential part of which involves the agent writing summary after every conversation. By automating this task, the average call handling time is decreased, leading to increased efficiency per agent.
• Seamless Context Transfer: Secondly, the summary proves invaluable when a call is transferred between agents. In scenarios where agents need to reconnect with a customer, they can easily pick up where they left off with the help of a call summary. Presenting each agent with a concise summary of the conversation thus far ensures that the context is seamlessly passed between agents and between calls, facilitating a smooth transition.

Wrapping Up

1. The Power of Contact Centers: Contact centers play a pivotal role in ensuring exceptional customer service across diverse businesses and industries.
2. Limitations of Deep-Learning/GenAI Models: While deep-learning models have been utilized for call summaries, they often fall short of meeting the specific needs of individual customers.
3. Introducing a Hybrid Solution: Our research proposes a hybrid system that not only offers out-of-the-box summaries but also integrates business-specific information through user queries using an intent-matching algorithm.
4. Customizable Summaries: With this innovative approach, businesses can tailor summaries to align with their unique objectives, fostering improved operational efficiency and ultimately boosting customer satisfaction.
5. Real-time Editing and Refinement: The streaming nature of our solution empowers agents to edit and refine generated summaries during calls, eliminating the need for time-consuming after-call work and improving overall productivity.

This work has been accomplished through a collective effort involving Aashraya Sachdeva, Sai Nishanth Padala, Anup Pattnaik, Varun Nathan, Cijo George, Ayush Kumar and Jithendra Vepa.

Learn more about how we’re changing conversation intelligence for contact centers around the world at Observe.AI.