Research Output
AspNet: Aspect Extraction by Bootstrapping Generalization and Propagation Using an Aspect Network
  Aspect-level opinion mining systems suffer from concept coverage problem due to the richness and ambiguity of natural language opinions. Aspects mentioned by review authors can be expressed in various forms, resulting in a potentially large number of missing or incomplete aspects. This work proposes a novel unsupervised method to extract aspects from raw reviews with a broader coverage. Previous research has shown that unsupervised methods based on dependency relations are promising for opinion target extraction (OTE). In this work, we introduce Aspect Network (AspNet), an AspNet that further improves existing OTE methods by providing a new framework for modeling aspects. AspNet represents the general indecomposable atom aspects and their dependency relations in a two-layered, directed, weighted graph, based on which the specific decomposable compound aspects in reviews can be effectively extracted. AspNet is constructed through an unsupervised learning method that starts from a small number of human-defined, domain-dependent aspects, and bootstraps generalization and propagation in a large volume of raw reviews. In summary, the major contributions of this work are twofold: Firstly, the proposed AspNet is a new framework in modeling aspects; secondly, an unsupervised method is proposed to construct AspNet in a bootstrapping manner within raw reviews to learn aspects automatically. Experimental results demonstrate that our proposed OTE method, based on AspNet, can achieve significant gains over baseline methods.

  • Type:


  • Date:

    19 September 2014

  • Publication Status:


  • DOI:


  • ISSN:


  • Funders:

    Historic Funder (pre-Worktribe)


Xia, Y., Cambria, E., & Hussain, A. (2015). AspNet: Aspect Extraction by Bootstrapping Generalization and Propagation Using an Aspect Network. Cognitive Computation, 7(2), 241-253.



Aspect extraction; Opinion mining; Aspect Network; Unsupervised learning

Monthly Views:

Available Documents