A Novel MEC Framework for Extractive Summarization Using Semantic Role Graph and Semantic Matching

With the progressive development of internet technologies, people are exposed to huge amount of information in various formats every day, and need to deal with them anytime and anywhere through different terminal devices. Among which, text information usually costs too much time for length and redundancy. Hence, providing an efficient text summarization approach is quite necessary. However, such approaches usually depend on artificial intelligence techniques requiring greater computing power that the terminal devices cannot afford. Therefore, in this paper we propose a Mobile Edge Computing framework for text summarization, in which the summarizing requests submission and results display are done on terminal devices, while the computationally intensive summarization work is offloaded to mobile edge servers deployed with summarization model. The efficiency and accuracy of the results obtained based on this framework depend largely on the summarization model used. Hence, we also propose an extractive summarization model using semantic role graph (SRG) and semantic matching model (SMM). First, construct a SRG based on the semantic roles obtained by semantic role labeling; then generate candidate summaries by traversing the SRG and choose the best one using a SMM based on Bidirectional Encoder Representation Transformers. SRG guarantees conciseness and accuracy of the abstracts at syntactic level though reducing the extraction granularity to semantic roles. SMM further guarantees accuracy at semantic level based on the prior knowledge learned through pre-training. In a word, the proposed framework together with the extractive summarization model using SRG and SMM can realize an efficient and accurate text summarization approach. Experiments on both in-house and public datasets validate this conclusion through comparison with existing methods.