A Simple Yet Effective Corpus Construction Method for Chinese Sentence Compression

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6874–6883 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6874 A Simple Yet Effective Corpus Construction Method for Chinese Sentence Compression Yang Zhao∗, Hiroshi Kanayama∗, Issei Yoshida∗, Masayasu Muraoka∗, Akiko Aizawa† ∗IBM Research - Tokyo 19-21 Nihonbashi Hakozaki-cho, Chuo-ku, Tokyo 103-8510 Japan yangzhao@ibm.com, {hkana,issei,mmuraoka}@jp.ibm.com †National Institute of Informatics 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430 Japan aizawa@nii.ac.jp Abstract Deletion-based sentence compression in the English language has made significant progress over the past few decades. However, there is a lack of large-scale and high-quality parallel corpus (i.e., (sentence, compression) pairs) for the Chinese language to train an efficient compression system. To remedy this shortcoming, we present a dependency-tree-based method to construct a Chinese corpus with 151k pairs of sentences and compression based on Chinese language-specific characteristics. Subsequently, we trained both extractive and generative neural compression models using the constructed corpus. The experimental results show that our compression model can generate high-quality compressed sentences on both automatic and human evaluation metrics compared with the baselines. The results of the faithfulness evaluation also indicated that the Chinese compression model trained on our constructed corpus can produce more faithful compressed sentences. Furthermore, a dataset with 1,000 pairs of sentences and ground truth compression was manually created for automatic evaluation, which, we believe, will benefit future research on Chinese sentence compression. Keywords: Chinese Sentence Compression, Corpus Construction, Dependency Tree, Pre-trained Transformer 1. Introduction Deletion-based sentence compression aims to delete words from a single sentence to produce a shorter ver- sion, that is, a compressed sentence that remains read- able and faithful to the meaning of the source sentence. Despite the lack of paraphrasing or word order change, this technique has proven useful in many applications, such as compacting subtitles for high-rate speech (Van- deghinste and Pan, 2004), shortening lengthy product titles on online retail platforms (Wang et al., 2018), be- ing used as a pipeline step in multiple document sum- marization (Banerjee et al., 2015), and improving neu- ral machine translations (Li et al., 2020). Over the past two decades, English sentence compres- sion has made significant progress in terms of unsu- pervised and supervised compression systems (Filip- pova and Strube, 2008a; Filippova et al., 2015), corpora of various genres (Knight and Marcu, 2002; Clarke and Lapata, 2006; Filippova and Altun, 2013), and evaluation (Clarke and Lapata, 2008a; Filippova and Strube, 2008a; Filippova and Altun, 2013). In par- ticular, Filippova and Altun (2013) utilized a depen- dency tree-based method to construct a large-scale En- glish parallel corpus consisting of pairs of sentences and compressed sentences for the first time by lever- aging English-specific dependency tree transformation, thus paving the way for follow-up studies to train their machine-learning models. In contrast, there has been very little progress in sentence compression in Chinese. The main reason for this is the lack of a parallel corpus of sufficient size and quality to develop and evaluate the Chinese compression systems. Also, it is …

Content truncated to 500 words