|
发表于 2023-8-10 17:36:30
|
显示全部楼层
近来的机器人作诗系统应该都脱胎于
Chinese Poetry Generation with Recurrent Neural Networks 这篇论文。就公开的系统来说,前有“
微软律诗绝句”,后有题主分享的编诗姬。
算法而言,机器作诗无非是基于字符级RNN的各种变形,加之格律的约束。答主之前也稍微有玩过,贴两首自己训练的模型做出的七言诗:
用的算法只是两层的字符级RNN,看起来就可以像模像样的作诗了。
然而机器人作诗的问题,仍然是有“诗感”而无“诗意”。从题主贴出的这首诗来看,只能说乍一看像一首诗,但是诗歌想要表达什么?句与句之间有怎样的演进?终究不禁细想。RNN虽然能通过学习一些长程的字词搭配,保持一首诗风格上,或者“诗意”上的统一,但是仍然不能说学习到了真正的语义。
说到底,计算机能够像模像样地写诗,还是利用了诗歌本身的模糊性和普通人对诗的不熟悉。当计算机开始写作文、写故事的时候,这种语义上的缺点暴露无疑。
从去年的工作来看,最好的通用故事生成器应该是
Skip-Thought Vectors,通过训练两组RNN,去encode小说中的句子,再decode出上句和下句。在论文最后,他们给出一个模型生成的故事是这样的:
she grabbed my hand . “ come on . ” she fluttered her bag in the air . “ i think we ’re at your place . i can’t come get you . ” he locked himself back up . “ no . she will . ” kyrian shook his head . “ we met ... thatcongratulations ... said no . ” the sweat on their fingertips ’s deeper from what had done it all of his flesh harddid n’t fade . cassie tensed between her arms suddenly grasping him as her sudden her senses returned to itsbig form . her chin trembled softly as she felt something unreadable in her light . it was dark . my body shookas i lost what i knew and be betrayed and i realize just how it ended . it was n’t as if i did n’t open a vein . thiswas all my fault , damaged me . i should have told toby before i was screaming . i should ’ve told someone thatwas an accident . never helped it . how can i do this , to steal my baby ’s prints ? ” 仍然是一个有小说感,但是nonsensical的东西。
不过基于Skip-Thought Vectors的一个看图讲故事的demo还挺像回事的,题主可以关注
A recurrent neural network for generating little stories about images:
总的来说,计算机创作在近期来看,仍然是篇幅越短、含义越模糊(越需要人脑补),越能够糊弄住人。未来只能期待DL在更好的记忆和处理语义上取得突破。 |
本帖子中包含更多资源
您需要 登录 才可以下载或查看,没有账号?立即注册
×
|