NLP Research and Practice at Hulu: From Historical Milestones to Product Development

The author, a Hulu NLP research engineer, shares insights on what Hulu and its algorithm engineers think while you watch, structuring the talk into three parts that blend personal experience with technical overview.

Starting with a brief NLP history, the article highlights three landmark models of the past two decades—Bengio's NNLM (2003), Mikolov's Word2vec (2013), and Vaswani's Transformer (2017)—explaining how each unified feature representations and advanced the field.

Detailed explanations follow for NNLM, Word2vec, Transformer, and BERT, covering their architectures, training objectives, and why they became foundational for a wide range of natural‑language tasks.

The piece then contrasts academic research with product development, noting differences in problem definition, data cleanliness, model selection, evaluation metrics, and engineering practices, and emphasizes the need to adapt elegant research models to noisy, real‑world product constraints.

Two concrete Hulu projects are described: (1) News personalization, which defines a recommendation problem, uses NLP features such as titles and subtitles, and employs dynamic topic mapping to match news to trending topics; (2) Content embedding, which creates vector representations of video metadata and tags using techniques like node2vec, enabling similarity‑based recommendations and downstream click‑through‑rate prediction.

Finally, the article outlines Hulu's AI Platform—comprising infrastructure, machine‑learning data, and AI service layers—that streamlines data pipelines, model training, versioning, and deployment, thereby improving both team collaboration and individual engineering efficiency.

The author concludes with a brief biography and contact information.