機械詩の部屋

Machine Poetry Room

Box, Printer, Printing paper, PC, AI, Human (2020)
 

石橋 友也  新倉 健人
    Tomoya Ishibashi Kento Niikura  

詩を綴るAIと、詩を綴るAIの文体を模倣する人間

An AI that writes poetry and a human who mimics the writing style of the AI that writes poetry

「Input」と「Output」と記された黒い大きな箱と記入台がある。体験者は詩のお題(タイトル)となる言葉を紙に記入し、黒い箱に投入する。数分後に箱からプリントアウトされた詩が2枚出てくる。1枚は大量の日本語文章を学習したAIが自動生成したもの、もう1枚はAI特有の文体を模倣する特殊な訓練を受けた人間によるものであり、一見すると見分けがつかない。


作者らは以前から深層学習によって詩(に見える短文)を自動生成する実験を繰り返してきた。自動生成させる詩の理想的なアウトプットのラインを見極めるために、作者自身も詩を大量に読み、書きながら実験を繰り返した。特に、AI特有の論理の飛躍や語尾の連続といった特徴が、子供の書いた詩や知的障害を持つ人の書いた詩と共通項があることを発見し、それらの詩を参考にしながらエンジニアリングを行った。そのような過程を経て、AIが綴る詩と我々が書く詩は近似していった。


シンギュラリティの議論が盛んにされる中、将棋や囲碁の世界ではAIの指し手から棋士が学ぶといった逆転現象が起きている。 本作では詩作という古典的な文章表現において、人間がAIを模倣するという転倒した構造の導入を試みる。

There is a large black box with "Input" and "Output" written on it and a writing stand. The participant writes the title of the poem to be composed on a piece of paper and inserts it into the black box. After a few minutes, two poems are printed out from the box. One is automatically generated by an AI that has been trained on a large volume of Japanese text, while the other is composed by a human who has undergone special training in mimicking the unique writing style of the AI. It is impossible to tell which is which at a glance.

We have made numerous experiments in generating poems (or short sentences that look like poems) using deep learning. In order to determine the line that defines ideal output for an automatically generated poem, we ourselves read and composed numerous poems while carrying out our experiments. In particular, we discovered that traits such as jumps in reasoning and repetition in word endings that are characteristic of AI have commonalities with poems composed by children or mentally handicapped people, and used such poems as references while engineering. By a process like this, the poems written by the AI and the poems written by us gradually became more similar.

Among widespread talk of the singularity, there is now a reverse phenomenon in domains such as shogi or go in which professional players learn from the moves suggested by AI. In this work, we attempt to create an inverted structure in which humans imitate AI in the traditional form of written expression known as poetry.

Photo: Ken Kato Photo courtesy: Tokyo Arts and Space

Photo: Ken Kato Photo courtesy: Tokyo Arts and Space

©2020 Tomoya Ishibashi