How AI Powers K‑12 Education: Insights from a Chief Algorithm Expert

AI Empowering Business Scenarios

Song Yang, chief algorithm expert at Zuoyebang, began his career in search and data mining at Baidu before focusing on question‑bank construction, live‑classroom AI, and various NLP and speech technologies such as translation, essay grading, text classification, intelligent tagging, and speech recognition, evaluation, and synthesis.

For young learners who cannot type, Zuoyebang developed a one‑click voice input for live‑classroom "voice bullet comments," improving recognition of unclear speech and limited context by enhancing short‑text speech models and incorporating domain‑specific language models.

In NLP, the team built a photo‑translation feature for K‑12 reading material, handling special structures like blanks and numbered lines, and applied NLP to knowledge‑graph construction for massive question‑bank tagging, achieving over 80‑90% coverage with semi‑automatic multi‑label tagging.

Voice and NLP are combined in quality‑inspection pipelines: speech is transcribed, then NLP detects potential issues using keyword‑based pre‑screening, enabling low‑precision but high‑recall automated checks that reduce manual effort dramatically.

AI Expectations: Gaps Do Not Mean Useless

AI often falls short of teacher expectations, especially in essay grading, but it still provides useful preliminary assessments for parents and helps teachers focus on higher‑level feedback.

Objective questions can already be fully automated, while subjective questions see partial AI substitution, with ongoing improvements.

Future: AI Will Evolve Quietly

Smart voice applications may not explode in popularity, but incremental improvements such as low‑latency, natural‑sounding text‑to‑speech for reading questions will become commonplace.

Large Models: Worth Trying in Speech

Pre‑trained large models, originally successful in NLP, are now being explored for speech, offering better base performance and reduced data requirements for domain adaptation.

Zuoyebang plans to leverage idle GPU resources at night for distributed training of such models.

End‑to‑End and Multimodal: Hype vs. Reality

End‑to‑end models have become the default at Zuoyebang, offering modest gains over traditional pipelines, while multimodal research (text‑image) shows limited short‑term impact on speech tasks.

JAX vs. Other Frameworks

JAX offers slightly better usability than TensorFlow but has not reached PyTorch’s popularity; Zuoyebang primarily uses PyTorch for its flexibility and community support.

Open‑Source AI

Open‑source frameworks accelerate AI development by reducing duplication of effort, though Chinese projects often lack extensive documentation compared to overseas counterparts.

AI Middle Platform vs. Data Middle Platform

AI middle platforms abstract common capabilities (e.g., OCR, speech, NLP) to serve multiple business lines, but their usefulness depends on company size and stage; overly rigid middle platforms can hinder rapid product iteration.

Overall, Song emphasizes that AI should serve concrete business needs, be pragmatically integrated, and evolve according to the organization’s maturity rather than chasing buzzwords.