A Discriminative Sentence Compression Method as Combinatorial Optimization Problem

Abstract

In the study of automatic summarization, the main research topic was `important sentence extraction' but nowadays `sentence compression' is a hot research topic. Conventional sentence compression methods usually transform a given sentence into a parse tree or a dependency tree, and modify them to get a shorter sentence. However, this method is sometimes too rigid. In this paper, we regard sentence compression as an combinatorial optimization problem that extracts an optimal subsequence of words. Hori et al. also proposed a similar method, but they used only a small number of features and their weights were tuned by hand. We introduce a large number of features such as part-of-speech bigrams and word position in the sentence. Furthermore, we train the system by discriminative learning. According to our experiments, our method obtained better score than other methods with statistical significance.