Localized Cebuano-English DRRM Chatbot via Retrieval-Augmented Generation and Large Language Models

Abstract

The study created and assessed a localized Cebuano-English Disaster Risk Reduction and Management chatbot using Retrieval-Augmented Generation and Large Language Models. It responded to residents' demand for clearer disaster communication, especially on typhoons, floods, evacuation instructions, emergency hotlines, go-bag preparation, and weather warnings. A developmental-descriptive research design was adopted in the investigation. The developmental aspect consisted of the design and deployment of a web-based chatbot prototype. The descriptive aspect evaluated current DRRM communication methods and user approval using the Technology Acceptance Model.

The system was built using an embedding-based RAG architecture. Knowledge entries from DRRM were sorted into structured knowledge chunks and stored in a repository, where they were transformed into embeddings and retrieved using vector-based search. The LLM produced localized Cebuano–English responses using only the approved retrieved knowledge. Also, the system has manual ingestion of Advisories and review by Admin/MDRRMO to guarantee only validated content is used by the chatbot.

Results showed that the top hazards of the respondents are typhoons and floods. Most residents’ sources of DRRM information are radio, barangay announcements, MDRRMO/LGU alerts, and social media. The results also revealed that the respondents favored Cebuano or Cebuano-English explanations, particularly for technical terms related to disasters. The TAM evaluation demonstrated a very high initial acceptability among users, with an overall grand mean of 4.45. This shows that the respondents regarded the chatbot as useful, easy to use, and acceptable as an additional DRRM communication support tool. The study said the chatbot can be useful for municipal and barangay DRRM communication, but it should not replace official advisories, evacuation orders, or emergency decisions given by authorized agencies.