Rule-based information extraction for mechanical-electrical-plumbing-specific semantic web

Information extraction (IE), which aims to retrieve meaningful information from plain text, has been widely studied in general and professional domains to support downstream applications. However, due to the lack of labeled data and the complexity of professional mechanical, electrical and plumbing (MEP) information, it is challenging to apply current common deep learning IE methods to the MEP domain. To solve this problem, this paper proposes a rule-based approach for MEP IE task, including a "snowball" strategy to collect large-scale MEP corpora, a suffix-based matching algorithm on text segments for named entity recognition (NER), and a dependency-path-based matching algorithm on dependency tree for relationship extraction (RE). 2 ideas called "meta linking" and "path filtering" for RE are proposed as well, to discover the out-of-pattern entities/relationships as many as possible. To verify the feasibility of the proposed approach, 65 MB MEP corpora have been collected as input of the proposed approach and an MEP semantic web which consists of 15,978 entities and 65,110 relationship triples established, with an accuracy of 81% to entities and 75% to relationship triples, respectively. A comparison experiment between classical deep learning models and the proposed rule-based approach was carried out, illustrating that the performance of our method is 37% and 49% better than the selected deep learning NER and RE models, respectively, in the aspect of extraction precision.