Identifying Call Center Topics with Natural Language Processing

Challenges

• A large financial company handled thousands of daily customer interactions with no systematic understanding of what customers were calling about.
• Decisions on product improvements and retention were made on intuition and manual spot-checks — leaving the business without a reliable foundation.
• The business needed a data-driven view of contact drivers and wanted to understand how call behaviour connected to the customers most likely to churn.
• Working with Danish-language call data added difficulty — most off-the-shelf NLP tooling is built and benchmarked on English, with limited support for Danish and insurance-specific terminology.
• Within the confines of GDPR, this also required strong anonymization techniques.

Approach

• Transcribed call recordings from audio to text to enable NLP analysis.
• Added anonymization algorithm to anonymize calls and filter out sensitive information.
• Integrated with customer and churn data.
• Developed custom tokenization to optimize richness and speed.
• Developed topic modelling (LDA) to identify topics that customers called about.
• Developed named-entity recognition (NER) to identify things and persons.
• Developed sentiment analysis to classify positive or negative topics.

Results

• End-to-end pipeline with transcription, anonymization, processing, NLP analysis and insights.
• Increased and systematic visibility into call center issues and performance.

Technologies & methods

Natural Language Processing (NLP) Sentiment Analysis Topic Modelling (LDA) Named-Entity Recognition (NER) Tokenization Speech-to-Text AI