A Hybrid Semantic Networks Construction Framework for Engineering Design

Abstract

In this paper, a novel approach is introduced for building semantic networks from scientific publication abstracts. With a seed statement as input, the approach generates a semantic network using Recursive Object Modeling (ROM), Skip-gram language modelling, and semantic similarity methods. Semantic Scholar API was used to retrieve data for building the ROM-based Semantic Networks (RomNet) following Environment-Based Design (EBD) methodology. The RomNet is then applied to an aircraft braking system design. The work includes two major contributions: a ROM-based phrase extractor and a Skip-gram model trained on automatically collected publication abstracts data. The phrase extractor was compared with two existing off-the-shelf key phrase extraction algorithms, namely TextRank and Rake. The ROM-based phrase extractor is capable of extracting most key phrases from target domains and shows higher precision, recall, and F-1 scores than other methods. The Skip-gram language model was evaluated using NASA thesaurus. We randomly sampled 457 pairs of similar domain-specific terms, which are related to aircraft braking and landing knowledge. We compared our Skip-gram model with Google's pre-trained word2vec model, as well as a baseline word2vec model. The experiment revealed that our language model was able to detect the most pairs of concepts from the NASA thesaurus. For the application part, the generated semantic network can be used for various down-streaming tasks, such as design information retrieval, computer-aided design idea generation, cross-domain communication support system, and designer training tool.