Metaphor generation based on noval evaluation method

doi:10.1016/j.neucom.2024.128651

2025-01-01

PRE

DOI：10.1016/j.neucom.2024.128651

原文链接

原文求助

摘要

En 中文

Metaphor generation is a difficult research area to study. In this task, the generated content must maintain certain elements of the original content, such as verbs, adjectives, and compound phrases that convey a deeper meaning than their literal meanings. However, there is a lack of a large and accurate parallel corpus for metaphor generation, which makes it difficult to train effective models. As a result, most current studies on metaphor generation rely on constructing pseudo-parallel corpora, which are often of low quality and lack diversity. To address this problem, we propose the creation of a large-scale annotated parallel corpus for metaphor generation. This corpus consists of 18,000 samples that have been carefully screened by professionals and cover metaphorical sentences that include verbs, nouns, and compound phrases. In addition, we propose an automatic evaluation metric called the MS-score, which is based on metaphoricity and similarity, to assess the quality of metaphors generated by the model. We validate the consistency of the automatic evaluation metric with manual evaluation metrics and use the MS-score to re-rank the generated metaphors.

Keyword:

Metaphor generation

Metaphorical encoding model

Parallel corpus

MS-score evaluation