Abstract: Most of the concrete dam construction management (CDCM) knowledge is recorded and stored in text form, which features a huge amount of data, severe fragmentation, and poor hierarchy. An important issue for the management is to mine construction knowledge from unstructured text data, clarify the logical relationship of knowledge, and improve the efficiency of knowledge application. This paper presents an intelligent method for generating CDCM knowledge graphs, or converting massive text data into directly usable knowledge. We combine word vectors, char vectors, a Bi-directional Long Short-Term Memory-Conditional Random Field (BiLSTM-CRF) network, and the attention mechanism; and build an intelligent model for CDCM entity recognition. It strengthens construction entity characteristics and extracts entity words from the management texts. Relationship types between entities are defined by the construction entities identified; mutual information is used to extract entity relationships and obtain construction knowledge chains; a CDCM knowledge graph is established by combining these chains. Application to a real management case shows this model gives an F1 value of 92.48%, outperforming other entity recognition models. Thus, this study demonstrates that the knowledge graphs can be established using the relationships between the construction entities recognized and that based on the graphs, a retrieval mechanism of the construction knowledge can be worked out for its rapid extraction and high application efficiency.

Key words: concrete dam construction management, knowledge graph, BiLSTM, attention mechanism, mutual information