The spring of artificial intelligence creation has come, and the "hidden rules" of ancient poetry are taught by computer.

  Xinhua News Agency, Beijing, March 12th Special feature: The spring of artificial intelligence creation is coming.

  Xinhua News Agency reporter Peng Qian Liu Si

  "Early spring rain Chu Qing, willow filar silk caught the shore warbler. The painting boat is full of smoke and waves, and the small bridge is light. "

  Who would have thought that this is a poem created by artificial intelligence with "early spring" as the key word. The author "Nine Songs" was developed by a student team led by Professor Sun Maosong of Tsinghua University Computer Science and Technology Department for three years.

  Xiao Bing, Microsoft’s chat robot, "narrowly beat" human opponents with a Chinese-style song "Peach Blossom Dream" in the variety show "Wit". When the singer sings such an archaic sentence as "In the boundless night rain, the past is like the wind, the peach blossoms in the ear laugh at the spring breeze, and you and I meet in the dream", many viewers think that this is made by human beings.

  The spring of writing poetry, composing music, painting and artificial intelligence has come quietly … …

  Technically, "familiar with 300 Tang poems"

  Zhou Ming, vice president of Microsoft Asia Research Institute, is a literary lover. He and the research team taught Xiao Bing to write modern poetry, and composing music is a new skill.

  Zhou Ming said that neural machine translation, chatting robot, reading comprehension and creation are the four main aspects of the application of natural language processing technology, and the difficulty is gradually increasing. Creation is at the top of the pyramid and it is the most difficult to break through.

  "In the beginning, people couldn’t find how to model creation because it was emotional. Writing poems and writing lyrics are all things with flying literary talents and sudden inspiration, which are difficult to capture. " He said. Thanks to the progress of deep learning and neural network technology in recent years, artificial intelligence creation has won many brilliant achievements.

  Zhou Ming introduced that writing songs by Xiao Bing is a process of "encoding and decoding". The researchers first trained artificial intelligence robots with popular songs. After sufficient training, you can start writing: after inputting the song theme in the form of keywords, it will be coded into a language that artificial intelligence can understand, and the robot will decode and output it in a way that people can understand, becoming a lyric; Then combine this word with the original keyword as a new input, and you can get the second sentence, and so on to get the whole song word. Similarly, you can also input lyrics and translate them into music scores, that is, you can regard music scores as natural languages, thus completing computer lyrics and music scores.

  "Nine Songs" learning to write poetry is also "familiar with 300 Tang poems, and you can recite them even if you can’t write poetry".

  Yi Xiaoyuan, the founder of "Nine Songs" and a graduate student in Tsinghua University, has input more than 300,000 ancient poems since the Tang Dynasty as a corpus, and used the deep learning model to make computer learning. In addition to the rules for leveling and rhyming poems, no artificial rules are given, but computers are allowed to learn the "hidden rules" in ancient poems themselves.

  "We don’t know the rules of how computers make such poems," said Sun Maosong, which is a "black box" phenomenon of deep learning. In his view, each ancient poem is like a necklace, and the beads on the necklace are words. In the deep learning model, the necklace is completely broken up, and then the implicit association between each bead and other beads is given different weights through automatic learning. When writing a poem, put different beads back into a new necklace.

  The ancients wrote poems mostly to express their feelings, and the style was sad, which also made the poems written by "Nine Songs" somewhat "sad for spring and autumn". The team hopes to make the "Nine Songs" more positive by strengthening the training of some relaxed emotional samples. In addition, how to write a longer poem on the basis of ensuring the consistency of the whole poem is also a new challenge.

  In artistic conception, "Kung Fu is beyond poetry"

  In addition to writing poems and composing music, artificial intelligence has evolved new skills such as writing novels and painting. The science fiction novel "The Day when a Computer Writes a Novel" created by artificial intelligence developed in Japan deceived all human judges and was successfully shortlisted for the Japanese Fiction Literature Award. Google artificial intelligence can also create paintings, and some paintings have been sold at a high price of $8,000.

  Artificial intelligence has defeated human beings in the fields of chess and cards, video games, etc. Does the latest progress in the field of artistic creation mean that it is not far from surpassing human beings in this respect?

  Zhou Ming believes that at present, artificial intelligence creation is only based on the imitation of big data, which is far from the creative intelligence of real human beings. "There is not enough data for training artificial intelligence creation, such as the data for composing lyrics; Secondly, the inspiration is not enough, because the data alone is just a follower, writing words and composing music are familiar, and there is no feeling of sudden inspiration in generate. "

  In Sun Maosong’s view, at present, artificial intelligence creation is a limited creativity, which theoretically does not exceed the creative space unconsciously defined by predecessors in the practice of poetry creation for thousands of years. The ancients wrote poetry as "Kung Fu is beyond poetry", often based on experience, with content and artistic conception, but it is difficult for machines to "express their aspirations" or "express their feelings through scenery" for the time being.

  Experts believe that the fear that artificial intelligence surpasses human beings in the field of creation is alarmist. "In areas that need deep connotation or inspiration, such as music creation, poetry and prose, I personally think that machines are basically unlikely to reach the state of people, but they will assist people’s creation." Zhou Ming said.

  However, artificial intelligence can improve the efficiency of professional creators. "If you can’t think of using a word when writing, artificial intelligence will think of a word for you and let you find that you can write like this." For ordinary people, artificial intelligence creation systems such as "Nine Songs" can lower the threshold of creation and realize the dream that "everyone can be a poet, painter or musician".

  Sun Maosong said that the team plans to upgrade the "Nine Songs" system in the future, so that it can judge the quality of people’s poems, such as whether there are mistakes in rhyme and level tone, whether words are unsatisfactory or sentences are incomplete, and help people improve their creative skills; You can also conduct "humanistic calculation" and quantitative analysis and research on ancient documents through big data.

  In addition to the creation itself, the study of machine creation can also bring unexpected benefits to other fields of artificial intelligence, and open up innovation in many fields and interdisciplinary subjects.

  "Before letting Xiao Bing learn music creation, we never thought that music can also be processed by natural language technology," Zhou Ming said. "So once the natural language breaks through, it can drive the breakthrough of cognitive intelligence and the whole artificial intelligence, and there will be many scenes that were unimaginable before."