news

original title: artificial intelligence may open up new poetry styles

【aspects・artificial intelligence poetry creation】

discussions on ai poetry in the humanities circle often revolve around "can ai write poetry?" in fact, this question is not difficult to answer. first of all, if poetry is defined as the expression of true feelings, then the current ai is indeed unable to write poetry. because true feelings mean having a mind, such "strong artificial intelligence" has not yet appeared. however, even if "strong artificial intelligence" appears, it may not write poems that are understandable by humans. because "strong artificial intelligence" is a completely different intelligence than humans, it may not even have individual consciousness. as far as the current ai is concerned, its form of existence is a whole rather than a group-according to the von neumann structure, the memory assigns parameters to the chip for calculation, and then receives the returned results. different ais have different algorithm mechanisms. if they can all write poetry, the difference will be far greater than poetry in different languages ​​in the human world.

so, can ai write poetry? the answer is yes. it is entirely possible for ai to write poems that look real and fake. in fact, ai poetry is more likely to pass the turing test (the test of whether it is human language) than the everyday language in which it was created. because the latter must take into account real-life experience, it has very high requirements in terms of fidelity. ai has no body and cannot grasp the sequence of human actions. it will inevitably reveal flaws when narrating. "embodied cognition" is still a problem faced by ai natural language. poetry can get around this because poetry is a special genre whose purpose is primarily aesthetic rather than practical. it can use fragmented and unfamiliar writing methods, such as "cuiping wets the sky beam and flies away in the middle of the night. the geese rise in the blue sky, and the few lines of writing are like old hiding places." "let her sit where she usually sits in the mirror and look out the window. as long as it is difficult for us to use daily linear time and causal logic to understand classic poems such as plum blossoms falling all over nanshan when we think of things we regret in life. this creative feature precisely eliminates the obstacles faced by ai.

the challenge before us is, how to improve the quality of ai poetry? since 2018, the pre-trained language models bert and gpt have made major breakthroughs based on the transformer algorithm. simply put, the principle is to convert natural language into vectors, and imitate the text in the database based on the probability of relationships between words. herein lies the problem of text refinement. if the number of texts is too large, such as using all tang poems as a database, which contains a large number of different or even conflicting styles, the ai ​​will be pulled in different directions during the learning process, and the poems written will be mediocre and mixed - using text in terminology, it means "lack of face".

to overcome this, the ai ​​needs to be trained in a style. developers should be familiar with poetry and carefully select poems with similar styles as materials for ai learning. it is not enough to divide based on the era. as qian zhongshu said, "tang poetry and song poetry are not just different in dynasty, but also in physical character." there are people in the tang dynasty who started the song tune, and there are also people in the song dynasty who inherited the tang tone. however, developers need to conduct more detailed screening. taking an old-style poetry ai currently on the market as an example, the developers used the "tongguangti" popular in the late qing dynasty for special training and achieved good results. for example, three five-rhymed poems are generated based on the first line "the night air is dimly white" in "a boat at night in wucheng" by the poet chen sanli:

the night air is white and the lights are red. the sorrow of traveling is broken by wine, and the dream of returning to poetry is poor. the road is broken, there are few people walking, and the words of the herons in the cold sichuan are empty. i groaned and cried bitterly, and cut my blood to answer the question of god.

the night air is white and the cold clouds are slowly lightening. the old love is still thin and drunk, and the dark moon has deepened. books that violate the customs often roll up, and dreams of cherishing people cannot come true. the thin shadow sings in depression, and i feel the autumn life when i raise my tent.

the night air is white and the setting sun is dimly red. the city is surrounded by mountains and the sun is shining down, and the water and wind are coming from the trees. outside the fragrant grass king and grandson, a guest in a lonely boat is thinking about it. i can't sleep while chanting, burning candles and spinning books are empty.

in these poems, except for a few places where words are made up ("cut blood to answer god") and go against daily experience ("night" is followed by "setting sun"), the three poems generally seem to have rich artistic conceptions, "traveling sorrow, fate and wine are broken" "returning to dreams is a poor poem" and "old thoughts are still drunk, and the dark moon has deepened" and so on are also quite amazing. compare chen sanli's original work - "the night air is white, and the smoke is green. the moon is cold in the lonely tent, and the stars are moving steadily on the gentle waves. the lights are noisy at the fishing port, and the vicissitudes of life change the sobriety. still thinking about the strategy, listening to the corner and looking at the lake pavilion "these poems are very close to it, fully demonstrating the effectiveness of style training.

poetry creation has always been about "thinking in the heart and speaking as poetry." so, what is the significance of ai poetry writing to human beings?

in the usual sense of auxiliary functions, ai has evolved to a more mature stage. the large ai model mentioned above has verified the conclusions such as "the seven sons of the ming dynasty learned from the tang dynasty" and "the same light body learned from the song dynasty". this means that literary styles that were previously considered abstract can be transformed into concrete data parameters. more importantly, for poetry, the reorganization of words often produces unexpected results. perhaps some sentences generated by ai can inspire poets to create, and even become auspicious elements in the development of poetry. long before the emergence of ai, there were attempts to break the limitations of human intentions to create poetry. in the early 20th century, parisian surrealists designed a game called "the exquisite corpse." without knowing each other, the participants randomly wrote words according to their parts of speech and finally combined them into sentences. the first sentence they got through this method was "le cadavre exquis boira le vin nouveau" (le cadavre exquis boira le vin nouveau). "le cadavre" means "drinking from the empty bottle" in french. the pun is quite clever.

in fact, poetry itself contains a tendency to transcend human experience. russian linguist roman jacobson believes that the so-called "poeticity" means that words point to themselves rather than reality. german poetry critic hugo friedrich believes that poetry "escapes the middle state of human beings, deviates from the usual objects and common emotions, abandons limited comprehensibility, and replaces it with ambiguous hints, in order to make people understand." poetry becomes an independent, self-directed construct whose content depends only on its language, its unrestrained imagination, or its unrealistic fantasy games, rather than on some kind of imitation of the world. a certain expression of emotion.” all these are in line with the characteristics of ai writing. from this dimension, ai may develop new poetry types and styles in the future.

(author: cheng yuhei, a full-time researcher at the institute of humanities and arts, shanghai jiao tong university)