Verb-Driven Task Networks Turn Static Epidemic Plans into Executable Playbooks

Public health emergency plans are meant to tell responders exactly what to do when a major infectious disease strikes, yet a new study argues that most of these documents fail at precisely the moment they matter most. Researchers in China have built a framework that converts the flat, often repetitive language of emergency plans into structured, executable task networks, and a case study suggests the resulting plans score far higher on completeness and operability than the originals they replace.

The work, published in Frontiers in Public Health, was carried out by Fangzhou Li of Zhengzhou Shengda University and Ming Cong, Yue Feng and Lili Rong of Dalian University of Technology. Their starting point is a diagnosis of why formal plans underperform: the authors describe existing documents as suffering from incomplete content, rigid structures, mechanical replication and insufficient operational feasibility, with plans at different administrative levels showing a high degree of homogeneity and little adjustment to local or hierarchical needs.

To quantify that homogeneity, the team compared the distribution of emergency-response verbs across plan texts and across records of real response activity. Cosine similarity between national, provincial and municipal plans ranged from 0.70 to 0.83, indicating that plans written at different levels of government closely resemble one another. Similarity between response records and the plans was markedly lower, at 0.46 against national plans, 0.52 against provincial plans and 0.55 against municipal plans. In other words, the documents that are supposed to guide action look more like each other than like the actions themselves.

The core insight of the framework is linguistic. In emergency response, operational actions are typically expressed in a verb-centered way, so the types and distributions of verbs largely reflect the task space a plan covers and its granularity. The researchers treat emergency-response verbs as the fundamental carriers of task semantics, using them to link the abstract language of plan documents with the concrete language of actual response records within a single modeling scheme built on Hierarchical Task Network (HTN) planning, a classical approach that iteratively decomposes complex tasks into subtasks and, ultimately, directly executable primitive tasks.

The resulting HTN-Verb framework assigns verbs to three hierarchical levels. Target verbs such as control, recover or respond correspond to complex, strategic tasks usually articulated by top-level authorities. Intermediate verbs such as organize or schedule correspond to coordinating subtasks associated with mid-level administrative entities. Operational verbs such as sample correspond to primitive tasks with urgency, measurability and direct operability, typically carried out by grassroots actors. To organize verbs into these levels, the team encoded them with the bert-base-chinese model, applied average pooling to obtain fixed-dimensional embeddings, reduced dimensionality with principal component analysis, and used agglomerative clustering with Euclidean distance and Ward linkage, fixing the number of clusters at three to match the task levels. Across 12 emergency themes the mean silhouette coefficient was 0.528, with theme-specific values between 0.490 and 0.544.

The evidence base came from 108 Chinese-language emergency plans, laws and regulations spanning national, provincial and municipal levels, supplemented by posts crawled from official Weibo accounts between January 2020 and December 2022 using keyword combinations pairing infectious disease or epidemic with emergency response. After segmentation with the THULAC toolkit, filtering for two-character verbs with a frequency of at least five, and manual screening, the researchers identified 579 verbs in response records and 177, 274 and 293 verbs in national, provincial and municipal plans respectively, yielding a task-oriented lexicon of 976 verbs that the authors have released publicly.

Because a single verb can carry different meanings depending on who performs it, the framework constrains verb interpretation by grounding it in the division of responsibilities within the emergency management system. The word assessment paired with a hospital refers to physicians’ clinical evaluation of patients, whereas the same verb paired with a municipal health commission denotes an administrative evaluation of epidemic trends. The team catalogued emergency-response agents, from the National Health Commission to subdistrict offices, and mapped 12 emergency themes, including material supply, medical treatment and rescue, transportation and logistics, and information reporting, to their primary responsible departments. A pilot annotation exercise with two independent coders reached 82.0 percent agreement on 50 verbs, with a Cohen’s kappa of 0.806.

To make task networks executable rather than merely descriptive, the framework models three types of horizontal relationships among tasks. Logical relations capture conditional dependencies using AND, OR and NOT structures; temporal relations describe precedence or simultaneity, such as screening before isolation; and functional relations capture dependency, independence or coupling between tasks, as when infectious disease surveillance and epidemiological investigation mutually shape each other’s outputs. When these relation types conflict within a single network, a priority rule resolves the tension: functional relations rank highest, followed by temporal and then logical relations, on the argument that disrupting critical functional dependencies would undermine response outcomes even if tasks run on time and remain logically consistent.

The dynamic generation mechanism closes the loop between planning and practice. As a response unfolds, newly emerging tasks not covered by existing plans are identified, recorded and folded into updated response records, enriching the verb lexicon and evolving the task structure through a cycle of plan generation, response execution, task augmentation and knowledge updating. The framework was evaluated in a grassroots case, Village A, an administrative village of roughly 500 households with a rural cooperative medical clinic staffed by two village doctors, which relies heavily on township and county authorities for expertise, supplies and funding. The village was selected because it offered both a formal emergency plan and records of actual epidemic-response activity, and the authors explicitly frame it as case-based evidence of applicability rather than a statistically representative sample.

The comparison is striking. The original plan text covered seven emergency themes with 98 emergency-response verbs, of which only 18 were operational-level verbs tied to 12 agents, whereas the actual response records involved all 12 themes, 201 verbs, 63 operational-level verbs and 21 agents. Using an evaluation system weighted by the Analytic Hierarchy Process, with plan adequacy measured by themes, agents, verbs and resource types and operability measured by task workflow connectivity and the share of operational-level verbs, the original plan scored 0.27 in overall effectiveness against 0.73 for the HTN-Verb-generated plan. Task workflow rose from 0.26 to 0.44 and the operational-verb share from 19.38 percent to 31.34 percent. The authors caution that the evaluation rests on a single case, that Weibo records may underrepresent informal frontline activity, that relationship identification still requires manual verification and lacks independent validation against annotated ground truth, and that no directly comparable computational method was benchmarked or real-world deployment tested. Those caveats notwithstanding, the study offers a concrete pathway for turning static plan documents into living task networks that can absorb what responders actually do.

Subject of Research: A verb-driven hierarchical task network framework for dynamically generating executable public health emergency plans for major infectious disease response

Article Title: A task-oriented framework for dynamic generation of public health emergency plans in major infectious disease response

Article References: Li, F., Cong, M., Feng, Y., & Rong, L. (2026). A task-oriented framework for dynamic generation of public health emergency plans in major infectious disease response. Frontiers in Public Health, 14, 1831748. https://doi.org/10.3389/fpubh.2026.1831748

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1831748

Keywords: public health emergency plans, hierarchical task network, HTN-Verb framework, infectious disease response, task decomposition, emergency management, verb lexicon, natural language processing, task relationship modeling, grassroots emergency response, Weibo response records, plan operability