结合排序学习的趣味成语生成模型

Model of Interesting Idioms Generation Based on Learning to Rank

汉语中的成语语言精练,典故源远流长,其趣味性是成语研究的一个重要方面.本文依据语音模板生成候选成语集合,从中提取语音、幽默、语义、情感和形态五大类11个特征,将这些特征融入到排序学习的相关算法中,从候选成语集合中检索趣味成语,进而构建趣味成语的生成模型.该模型将成语生成问题映射到信息检索领域,以查询及相关反馈的技术解决生成问题.经机器和人工的双重评估,实验结果表明五个维度的特征能够细致刻画趣味成语,生成质量较高,模型具有一定的实用价值.

Chinese idioms feature a concise form,most of which come from centuries-old allusions. Interestingness of idioms is an important aspect of idiom research. In this paper,we use phonetic templates to generate candidate sets of idioms,and extract eleven features of these idioms respectively from five dimensions,namely,phonetics,humor,semantics,sentiment and morphology. We integrate these features into the relevant algorithms of learning to rank so as to retrieve interesting idioms from the foregoing candidate set for an idiom generation model. The learned model transforms the problem of idiom generation into an information retrieval problem,and uses query and relevance feedback technology to solve the generation problem. Based on both manual and automatic evaluation,experimental results showthat interesting idioms can be precisely described and screened out based on these five-dimensional features. The idiom-generating model proves to have high quality and certain practical value.