Automatic Knowledge Extraction from Environmental Reports with Reference to Time and Spatial Coordinates of Water Bodies

The paper presents a new method for automatically extracting environmental knowledge from reports and news texts related to facts about the state of river waters or their pollution. Knowledge extraction is carried out taking into account the binding of the obtained facts to the spatial coordinates of specific water bodies and time intervals. The relevance of the work is due to the significant availability of such environmental data in the news, websites of institutions, and social media, and the need for their quick and accurate processing. The proposed method combines the detection of facts about the state of waters or their pollution, recognition of geographical names from the text and headlines, as well as the determination of time features by analyzing the hierarchical structure of the document. The method optimizes the contextual-semantic criterion, which maximizes the completeness and probability of detecting all existing connections between key phrases in the text of facts, time periods and water bodies and, at the same time, minimizes the number of false positive connections between them, by formalizing the connections in the form of “subject–predicate–object” (SPO) triplets and using the Jaccard measure to find the degree of similarity between the lists of key phrases that characterize these facts and water bodies. Knowledge extraction is based on identifying and using the hierarchical structure of the document, using large language models, and actualization the knowledge base with information with Retrieval-Augmented Generation (RAG) for regular knowledge update and binding to the time intervals and spatial coordinates. The result is a structured knowledge base in the form of “fact – water body – time interval” triplets, which can be used to analyze the dynamics of water status, identify trends, and make management decisions to improve the state of surface waters. The result of applying the proposed method is presented using the example of the annual report on the activities of the Southern Booh River Basin Water Resources Management for 2019, which illustrates its efficiency.