Analysis of Characteristics of Complaints on Parenting Q&A Sites Using pLSA and Data Augmentation

This study investigates the classification and clustering of complaints on a Japanese parenting Q&A site, aiming to identify meaningful patterns from limited labeled data. To address data scarcity, generative AI was utilized for data augmentation through prompts that reflected authentic parenting frustrations, with synthetic data validated by comparing classification performance under varying proportions of generated content. Complaint texts were vectorized using Bag-of-Words, Doc2Vec, and Sparse Composite Document Vectors, providing multiple levels of semantic representation. LightGBM was used as the classifier, and F1 scores measured performance. Clustering of predicted complaints employed probabilistic Latent Semantic Analysis, with topic numbers selected via Bayesian Information Criterion. Six distinct themes emerged, including childcare stress and family conflict. Incorporating generated data improved the F1 score from 0.824 to 0.865. The findings highlight the potential of generative AI to augment low-resource datasets and demonstrate the effectiveness of context-aware embeddings and probabilistic clustering in structuring real-world text data.