One of the most underrated NLP techniques in my experience is sentence embeddings combined with semantic similarity search. While transformers and large language models get most of the attention, embeddings are extremely practical for real-world applications.

Why they’re often overlooked:

• Cross-document retrieval: You can identify relevant information even when keywords don’t match exactly.
• Clustering and topic discovery: Embeddings allow grouping of similar sentences or documents and uncovering latent topics.
• Efficiency: Once embeddings are generated, similarity searches are computationally cheaper than running full model inference every time.
• Explainability: You can inspect nearest neighbors to understand why a particular result is returned, which is useful for analysis and trust.

In practice, sentence embeddings are versatile for building search engines, recommendation systems, summarization pipelines, and improving chatbot retrieval, all without requiring heavy model inference for every query.

Other underrated techniques I’ve seen include attention visualization for interpretability and clever domain-specific tokenization, but embeddings consistently provide value across a wide range of NLP applications.