DOFM: Domain Feature Miner for robust extractive summarization

Abstract

The domain feature retrieval has potential applications in text summarization. However, it is challenging to mine domain features from the user reviews. In this paper, a novel Domain Feature Miner (DOFM) is designed by (i) formulating the feature mining problem as a clustering problem and (ii) engaging three newly conceived empirical observations such as frequency count, grouping semantics, and distributional statistics of features. Later, Symmetric Cluster Extraction (SCE) and Asymmetric Cluster Extraction (ACE) algorithms are designed to identify domain features from clusters. The effectiveness of the DOFM is verified on benchmarks provided by the University of Illinois at Urbana–Champaign and compared with the four state-of-the-art (SOTA) approaches using Precision, Recall, and F-score. Moreover, ROUGE (Recall-Oriented Understudy for Gisting Evaluation), a well-known package for automatic evaluation of summaries is used to evaluate the DOFM generated summaries. The Error Analysis reveals that at least one of three annotators would prefer 84% sentences of all DOFM generated summaries, while 36% sentences are preferred by all three. This indicates the robustness of DOFM in domain feature retrieval and extractive summarization.