**Background:** Natural language processing (NLP) has advanced rapidly due to large-scale text data and transformer architectures, enabling tasks such as question answering, summarization, and translation at near-human levels. Despite this progress, systematic adoption of NLP in the humanitarian sector—the ecosystem of organizations providing assistance during crises—remains extremely sparse, with few initiatives scaling beyond the pilot stage. The authors argue that limited awareness of NLP's potential, lack of domain-specific datasets and standards, and ethical concerns around bias and accountability are key barriers. This paper aims to provide a pragmatically-minded primer to humanitarian NLP, identify concrete applications, and discuss challenges and solutions to foster an impact-driven community.
**Methods:** This is a narrative review and position paper. The authors first provide a high-level technical primer covering distributional semantics, transformer architectures (e.g., BERT, GPT), transfer learning, and limitations such as data hunger, bias, and black-box nature. They then describe the humanitarian ecosystem, including the cluster system, Humanitarian Needs Overviews (HNOs), and the Humanitarian Response Plan (HRP). The paper identifies three application areas: (1) tracking external data sources (social media, news) to monitor and predict crises, citing the 2010 Haiti earthquake crisis mapping via Ushahidi as a case study; (2) empowering communication with affected populations through chatbots, SMS-based needs assessment, and machine translation for low-resource languages, referencing Translators Without Borders' Gamayun project and Meta AI's 200-language translation model; (3) generating structured datasets from unstructured text using named entity recognition and relation extraction, and conversely using generative models (GPT family) to produce reports from structured data. The authors present the Data Entry and Exploration Platform (DEEP) as a collaborative analysis tool that has supported HNO production in Afghanistan, Somalia, South Sudan, and Sudan, and has been used for crises including Mozambique, Pakistan, Libya, Nigeria, Yemen, Ukraine, and Venezuela. They describe HumSet, a multilingual dataset of approximately 17,000 annotated documents in English, French, and Spanish, covering 46 global humanitarian response operations from 2018 to 2021, annotated using a unified humanitarian analytical framework with three label categories: Sectors, Pillars/Sub-pillars 1D, and Pillars/Sub-pillars 2D.
**Key Results:** The paper does not present original experimental results but summarizes initial evaluation studies from Fekih et al. (2022) showing that fine-tuning large pre-trained language models on HumSet text-label pairs yields promising benefits for extracting informative humanitarian text excerpts and classifying them according to the humanitarian analytical framework. The authors note that CrisisBench, a benchmark dataset of social media text labeled by relevance, intent, and sector, exists but that tools like AIDR and MicroMappers are no longer maintained. They report that DEEP has directly informed Flash Appeals, Emergency Plans for Refugees, Cluster Strategies, and HNOs in multiple countries. No quantitative performance metrics, p-values, or confidence intervals are reported.
**Clinical Implications:** This paper does not address clinical care directly. Its implications are for humanitarian response operations, including improved situational awareness through real-time social media monitoring, more inclusive needs assessment via multilingual and low-resource language technologies, and more efficient data analysis through automated information extraction. The authors emphasize that ethical deployment requires robust evaluation protocols to address bias, explainability, and data privacy, and that cross-functional collaboration between humanitarians and NLP experts is essential for safe and impactful applications.