How to Avoid Sentences Spelling Boring? Towards a Neural Approach to Unsupervised Metaphor Generation

Proceedings of NAACL-HLT 2019, pages 861–871 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 861 How to Avoid Sentences Spelling Boring? Towards a Neural Approach to Unsupervised Metaphor Generation Zhiwei Yu and Xiaojun Wan Center for Data Science, Peking University Institute of Computer Science and Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {yuzw,wanxiaojun}@pku.edu.cn Abstract Metaphor generation attempts to replicate hu- man creativity with language, which is an at- tractive but challengeable text generation task. Previous efforts mainly focus on template- based or rule-based methods and result in a lack of linguistic subtlety. In order to crea- te novel metaphors, we propose a neural ap- proach to metaphor generation and explore the shared inferential structure of a metaphorical usage and a literal usage of a verb. Our ap- proach does not require any manually annota- ted metaphors for training. We extract the me- taphorically used verbs with their metaphori- cal senses in an unsupervised way and train a neural language model from wiki corpus. Then we generate metaphors conveying the assigned metaphorical senses with an improved deco- ding algorithm. Automatic metrics and human evaluations demonstrate that our approach can generate metaphors with good readability and creativity. 1 Introduction Metaphor is a kind of language filled with vitali- ty and elasticity. It employs words in a way that deviates from their normal meaning to represent another concept (Li and Sporleder, 2010). Using metaphor allows us to express not just informa- tion but also real feelings and complex attitudes. There is a clear need in computational metaphor generation whose insightful results can be used in many applications from entertainment to educati- on (Veale, 2016). Besides, a unified metaphor an- notation procedure and creation of a large publicly available metaphor corpus are in great demands. Such resources make it possible to interpretate the identified metaphorical expressions and enhance the performance on other Natural Language Pro- cessing (NLP) applications (Shutova, 2010). Although metaphor has a long history of aca- demic studies in both philosophy and linguistics Figure 1: The Metaphor Generation Process (Genesereth, 1980; G. and M., 1985), it still re- mains a tough problem for the NLP communi- ty. As the metaphor is hardly to be well-defined and modeled, little work focuses on the metaphor generation. Most of the previous efforts rely on hand-coded knowledge (Martin, 1990; Feldman and Narayanan, 2004; Agerri et al., 2007) , which heavily constrains the diversity of generated meta- phors. The end-to-end approach presented to sequence learning has been proved effective on the gene- ration tasks like machine translation (Sutskever et al., 2014), abstractive summarization (Tan et al., 2017), product review generation (Zang and Wan, 2017) and multi-label classification (Yang et al., 2018). The approach is able to train a langua- ge model which can generate fluent and creative sentences with sufficient corpus. Unfortunately, in spite of the industrious exploration of the meta- phor corpus, the annotated data available now is still far from training a good language model. …

Content truncated to 500 words