A Hybrid PSO-GA for Extractive Text Summarization

A Hybrid PSO-GA for Extractive Text Summarization Bui Thi Mai Anh, Nguyen Thi Thu Trang∗, Tran Thi Dinh, Nguyen Thi Hoai Hanoi University of Science and Technology School of Information and Communication Technology anhbtm@soict.hust.edu.vn, trangntt@soict.hust.edu.vn dinh.tt173015@sis.hust.edu.vn, hoai.nt202195m@sis.hust.edu.vn Abstract In extractive summarization, searching for the most salient sentences to generate summaries can be resolved by applying meta-heuristic optimization algorithms including genetic algorithms (GAs), par- ticle swarm optimization (PSO), etc. In these approaches, the best solution, i.e. the expected summary, is found by optimizing an objective function, tak- ing into account the features of docu- ment and/or sentence. However, these traditional algorithms alone may suf- fer from a weak local search capability and slow convergence speed. To this end, this paper proposes a novel hybrid GA-PSO algorithm, namely PSOGA- BKSum, for single document text sum- marization based on a scoring feature subset and updating the candidate toward the best solution technique. The experi- ment results on the two common extrac- tive datasets, DUC2001 and DUC 2002, have shown that PSOGA-BKSum out- performs some state-of-the-art works on all three ROUGE point metrics. PSOGA- BKSum has considerably progressed on DUC2001 dataset, from 4.3 to 9.7% higher (i.e. 8.5 to 21.9% relatively) than the best ROUGE-1 score of previous un- ∗*Corresponding author supervised algorithms for this problem. For the dataset DUC2002, the PSOGA- BKSum receives the higher points on ROUGE-1 and ROUGE-2 scores, around 4.3% to 8.7% than the best previous al- gorithms. 1 Introduction The automatic text summarization allows users to reach information in a shorter form without losing any important aspects presented therein. This task becomes increasingly important due to the explosion of textual information on the Internet. Depending on the way in which sum- maries are generated, there are two approaches of text summarization (Hahn and Mani, 2000). Extractive summaries are formed by selecting from the original document the most salient sentences. Abstractive summaries, on the other hand, require advanced linguistic techniques to rephrase and generate new sentences which are not in the original document, therefore are much more complex than the former. The extractive approach can be studied as a binary classification task in which sen- tences from the input document are split into two groups: in-summary and not-in-summary. There is a huge diversity of techniques to au- tomatically generate extractive summaries for a single document which can be grouped into two directions: supervised and unsupervised. The former methods are based on machine learn- ing models (Wong et al., 2008) or deep learning models (Liu, 2019), those require a huge train- ing dataset including human-generated sum- maries, then very costly and time consuming. The latter approaches, on the other hand, do not require any training corpus but aim to rank all the sentences from the original document by exploring their relationship. The sentences with the highest ranking scores are then se- lected to build the summary. In this research direction, many techniques have been proposed for extractive summarization such as Hidden Markov Model (Yang et al., 2014), graph- based model (Mihalcea, …

Content truncated to 500 words