[1I3-GS-10f] AI application

2:10 PM - 2:25 PM JST(5:10 AM - 5:25 AM UTC)

[1I3-GS-10f-03] A Study on Multi-Label Distribution Estimation from Single-Label Customer Complaint Data

〇Masataka Suzuki¹, Tomoko Sasano¹, Ayako Yamagiwa¹, Kensuke Sato², Masayuki Goto¹ (1. Waseda University, 2. Meiji Yasuda Life Insurance Company)

Keywords:

Multi-label,Complaint Data,Multi-label Expansion,Single Positive Multi Label

In enterprises, complaint data analysis is essential for improving customer satisfaction. Some companies conduct statistical analysis using cause labels assigned to each complaint record by responders. However, the labeling system was established empirically, and there is no certainty that it is the optimal method, including the fact that a single label is assigned to each complaint.Previous studies attempted to build machine learning models to estimate cause labels from complaint descriptions but faced the challenge of accuracy remaining around 70%. One reason for this could be that the constraint of a single label does not reflect the true causal structure. Furthermore, a complaint texts often contain noise such as fixed phrases and operational communications, making them unsuitable for direct analysis.Therefore, this study proposes a multi-label classification method that extracts relevant sentences from complaint data as input information, treats the existing single label as a subset of the true label set, and estimates the missing additional labels. Validation experiments using real data will confirm the effectiveness of the proposed method by comparing it with human annotated data. Furthermore, through practical application examples, we will demonstrate that it is possible to grasp complex customer needs that could not be captured by traditional single-label analysis.