Active Learning and Model Enhancements for Semantic Tag Mining in 58.com Voice Data

The paper introduces the importance of voice as a communication medium on 58.com, where large volumes of voice data from call‑center, telephone, and micro‑chat scenarios provide valuable mining opportunities for semantic tagging.

Two main stages are described: converting voice to text using an in‑house ASR engine, followed by semantic tag extraction using natural language processing techniques such as keyword mining, text classification, and entity recognition.

To solve the cold‑start issue of limited labeled data, an active learning framework is applied. The iterative process involves training an initial model on a small labeled set, scoring unlabeled samples, selecting high‑value samples for annotation based on uncertainty or diversity queries, and retraining until the unlabeled pool is exhausted or performance stabilizes.

Uncertainty‑based query strategies (minimum confidence, margin sampling, entropy) and diversity‑based strategies (text similarity) are evaluated. Experiments on real estate semantic tags show that uncertainty sampling reduces required annotations by ~40% compared to random labeling, while diversity sampling saves ~60%.

Various models are benchmarked: keyword matching (baseline, 98.75% accuracy, 37.44% recall, F1 54.3%), XGBoost (F1 59.21%), TextCNN (F1 64.97%), a fused model combining the three (F1 70.15%), CRNN (leveraging TextCNN + RNN for long‑range context, F1 76.08%), and an improved Wide&Deep model that replaces the Deep component with TextCNN (F1 80.67%).

Online A/B tests confirm that two rounds of active learning (first uncertainty, then diversity) raise recall from 52% to 80% while maintaining 98% accuracy, achieving a 28% absolute recall improvement.

The study concludes that active learning dramatically cuts annotation costs and that the improved Wide&Deep architecture delivers the best performance; future work will explore pre‑trained language models and adversarial training to further boost results.