実践 自然言語処理 review: deploy and ops, not just models

The review frames its argument around a gap in NLP publishing: books tend either to dig into algorithm theory or to walk through library usage, while what actually causes pain in the field is how to put a model into production, grow it, and keep it running. 'Practical Natural Language Processing' is presented as filling that gap by treating the whole project lifecycle, from data collection through deployment and monitoring, as one unit.

The book's stated stance is to start simple: begin with rules, move to machine learning once data accumulates, and only then go to deep learning. The review singles out the explicit position that deep learning is not a silver bullet in industry settings, citing overfitting on small datasets, domain adaptation difficulty, low interpretability, cost, and on-device deployment constraints as reasons. Rules and heuristics, meanwhile, are still valued both for building a first version quickly and for filling holes in production systems.

The final chapter on the end-to-end NLP process is where the review spends the most energy. Deployment is broken into packaging, serving, and scaling; model size is treated concretely, with the note that a trained Word2vec model is about 4.8GB and takes over 100 seconds just to load into memory. Beyond deployment, the review points to the book's material on covariate shift, reproducibility, monitoring, technical debt, and the non-technical side of AI projects, including the observation, attributed in the book to Gartner research, that more than 85% of AI projects fail.