谷歌人工智能會寫詩 而且你一定能讀懂

After its attempts to digest romance novels, one of Google's artificial intelligence projects is now accidentally writing poetry, some of which would make the fictional Vogons proud.

在讀完幾本言情小說後，谷歌公司開發的一項人工智能項目於偶然間寫出了幾段詩歌。其中一些可能會令沃貢人對自己種族的詩歌感到驕傲（沃貢人是小說《銀河系漫遊指南》中虛構的種族，以詩歌質量低下聞名）。

Google is working with Stanford University and University of Massachusetts in the US to enhance the natural language skills of an AI technique called recurrent neural network language model (RNNLM), which is used within machine translation and image captioning among other tasks. It essentially builds sentences a single word at a time by analysing the previous words in that sentence.

谷歌目前正在和斯坦福大學及馬賽諸塞大學合作，改進一項名爲循環神經網絡語言模型的人工智能技術（RNNLM）。這項技術普遍應用於機器翻譯和圖像捕捉的任務中。而在人工智能搭建句子時，該技術會對句子中已有的單詞進行分析，從而生成每個新的單詞，在這一過程中起到至關重要的作用。

Currently, RNNLM is not capable of implementing global themes or features, such as a set topic, within its sentence generation. Each sentence produced by the algorithms doesn't necessarily flow smoothly into the next. The work, published as a paper through Cornell University's open scientific paper archive, arXiv, details the researcher's efforts to add the ability to apply a global theme to sentence generation using a system called a variational autoencoder and the results, including what could easily be mistaken for poetry.

目前循環神經網絡語言模型的組句部分尚不具備處理包括設定題目在內的全局主題和特徵的能力，而這種算法產生的句子在前後銜接得並不順暢。這項研究以論文的形式發表在康奈爾大學公開的論文庫arXiv中。論文詳細描述了研究者是如何運用名爲變分自動編碼器的系統幫助該算法實現識別全局主題的功能的，而得出的結果極易令人將其誤認成詩歌。

Some of it would arguably give Douglas Adams a run for his money, as the creator of the Vogons, the Hitchhiker's Guide to the Galaxy and its Ode to a Small Lump of Green Putty I Found in My Armpit One Midsummer Morning, during a recital of which “four of the audience members died of internal haemorrhaging and the president of the Mid-Galactic Arts Nobbling Council survived only by gnawing one of his own legs off”, according to the book.

道格拉斯·亞當斯創造的《銀河系漫遊指南》中一位詩人朗誦的《關於一個仲夏的清晨我在自己的腋下發現一團綠色膠泥的詠歎詩》糟糕到令“四位聽了詩歌的委員會成員死於體內大出血，而銀河中部藝術委員會的會長嚼完了自己的一條腿才保住了性命。”而如今人工智能寫出的句子則爛到可以與這首詩一較高下了。

The researchers fed the system starting and ending sentences and then asked it to fill in the gap. Below, the bolded text is what the researchers gave the algorithm and the text in between is what it produced. The results are generated by the machine based on what it learnt from thousands of romance novels, which made some of the topics “rather dramatic”, according to the authors of the paper.

研究者準備了一段話的開頭和結尾，要求系統進行填空。圖中加粗的內容是研究者提供的內容，而兩句話中間部分則是系統合成的內容。系統在閱讀了上千本言情小說後才生成了這些句子。論文作者表示，看了這些言情小說令系統寫出的一些句子“頗具戲劇性”。

The generated sentences make grammatical sense, maintain a sort of theme and for the most part fit with the start and end sentence. Others weren't quite as poetic, but still maintain the theme set by the start and ending sentences.

這些句子合乎語法規則，遵循了同一主題，大部分內容與開頭和結尾的句子相吻合。有些句子並未明顯體現出詩的感覺，卻仍能起到承上啓下的作用。

The results show interesting improvements in the ability of the machine to generate sentences that make sense together, which could led to much more human-like interactions with AI chatbots, perhaps even Google's Now.

上述結果反映了人工智能在生成有效句子方面富有趣味性的進步。這樣的進步能促進人類和人工智能聊天機器人進行更多類人互動，互動的對象甚至可以擴展到谷歌的語音助手Now。