Classifying Semantic Relationships of Utility-Specific Terminologies via LSTM Networks along Shortest Dependency Paths

The inconsistency of vocabulary terms used by different utility organizations makes data integration from distinct sources a challenging task. A semantic resource that classifies utility terminologies into sets of synonyms, hyponyms, and meronyms enables the computers to interpret the lexical semantics and to avoid mismatches when integrating heterogeneous utility data. However, constructing such semantic resources requires significant amounts of effort and time. This paper presents a shortest dependency path (SDP)-long short-term memory (LSTM) approach to automatically classify the semantic relations (i.e., is-a, part-of, is-similar, and random relationships) of two utility terms in a definition sentence. SDP is the shortest dependency path between two domain terms in a sentence, which retains the most informative information to relation classification while eliminating irrelevant words in the sentence. SDP-LSTM leverages LSTM units to pick up heterogeneous feature information along the SDPs and conducts automatic feature learning for semantic classification. The proposed approach was tested on a corpus of definition texts collected from utility design manuals. The preliminary results show an overall accuracy of over 80% and thus, the newly created method can serve as a good starting point to construct a semantic resource for the utility domain.