Text Summarization for Call Center Transcripts

While text summarization of transcripts in call centers is needed for detailed analysis, it presents challenges stemming from the call itself (context switching among speakers, cross talk, etc.) and from the resulting transcript (ASR transcription errors). This work aims to develop a summarization model suitable for on-premise deployment at call centers by fine-tuning pre-trained open-source large language models, assisted with reference summaries generated by GPT-3. The results are analyzed using ROUGE and human evaluation scores, and the correlation of these two metrics is examined. A fine-tuned BART model outputs satisfactory summaries with a human evaluation score of 6.95, approaching the GPT-3 score of 7.69.