Transformer Call Summarization for Noisy Transcript Noise Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing call summarization systems face challenges in generating accurate summaries, particularly when calls deviate from structured formats or contain 'noisy' information, leading to laborious manual processes and suboptimal automated results, which increase handling time and reduce productivity for customer service experts.
Innovation Solution
A transformer-based call summarization system that automatically generates summaries by processing call transcripts using natural language processing models, leveraging fine-tuned ML models to create situation and trouble descriptions without relying on structured frameworks, thereby improving noise handling and reducing computational costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated processes are used to generate call summaries, then productivity is improved and handling time is reduced, but the system fails when calls contain noisy information or deviate from structured formats
Solution Approach 1:
The patent transforms the input call data from unstructured or semi-structured formats into a standardized structured format through automated processing. This parameter transformation enables reliable extraction of key information (situation, trouble, action, resolution) even from noisy calls, resolving the contradiction between automation speed and accuracy with unstructured data
2Measurement precision
If manual summarization is performed by experts, then summary quality is maintained, but handling time increases and productivity decreases
Solution Approach 1:
The system enables self-service by automatically generating call summaries without requiring expert intervention. The automated process extracts and structures key call information independently, achieving both high productivity (summaries generated for all calls including those with noisy data) and maintained quality (structured STAR format outputs), thus resolving the contradiction between manual quality and automated productivity
3Loss of time
If existing automated summarization systems are used, then handling time is reduced, but computational costs are high and timeout errors increase
Solution Approach 1:
The patent extracts only the essential information needed for call summaries (situation, trouble, action, resolution elements) from call transcripts using targeted natural language processing. This selective extraction approach reduces computational overhead compared to processing entire transcripts, thereby reducing both computational costs and timeout errors while maintaining summary quality and speed
Data Source
AI summary
A processor may receive a call transcript including text and form a text string including at least a portion of the text. The processor may generate a situation description of the call transcript, which may comprise processing the text string using a transformer-based machine learning model. The processor may generate a trouble description of the call transcript, which may comprise creating a sentence embedding of the situation description, creating sentence embeddings for a plurality of utterances within the portion of the text, determining respective similarities between the sentence embedding of the situation description and each of the sentence embeddings for each respective one of the plurality of utterances, and selecting at least one of the plurality of utterances having at least one highest determined respective similarity as the trouble description. The processor may store a call summary comprising the situation description and the trouble description in a non-transitory memory.