This is a section of Shakespeare’s sonnets. This is a sonnet of Deep-speare’s Artificial Intelligence program “Creating.” For the time being, the quality of Deep-speare’s “work” is, at least in terms of rhythm, rhyme and grammar, and the performance of the 14 lines of this section is very good enough to confuse many people. That’s why some research teams have found that most readers can’t distinguish between artificial intelligence-generated poetry and human poetry.
The Deep-speare team, consisting of three machine learning researchers and a literary scholar, used about 2,700 14-line poems and about 367,000 words to train the artificial intelligence “poet” to learn to “create” itself.
Simply put, Deep-speare uses deep learning to filter the poems in the training database, trying again and again to create verses that match the sample.
Although in previous similar projects, researchers have given artificial intelligence advance knowledge of rhyme, rhythm, and so on, Deep-speare independently learned three elements associated with 14 lines of poetry writing: rhythm, rhyme, and natural language (i.e., words are correctly and smoothly combined).
Specifically, Deep-speare’s system consists of three parts: a rhythmic model, a rhyme model, and a natural language model that ensures proper grammar, of which the natural language model is the most important part.
First, the language model filters and predicts words in a corpus (the content of the corpus is based on Wikipedia terms, Reddit topics, and some databases that are built specifically) to determine which words are suitable for combination into sentences. After proper training, the language model gives fluent sentences high scores and meaningless sentences low.
The quality of the language model can be extracted by observing the association intensity of the next word (right-to-left). For example, “San-Francisco” often appears at the same time, and “coffee” is often more relevant to words such as “refresh” or “life-giving” than to words such as “powerful” or “light”. If the language model can handle this information correctly, it can be assumed that the model has largely captured the complexity of the language.
Once the language model is well-trained, it is no longer difficult to generate a sentence from scratch, and repeating this step can realize the basis for creating fourteen lines of poetry.
In addition to words and sentences, Deep Speare learns the rhythm of observing the letters and punctuation in each line and determining which characters correspond to which syllables and which syllables receive accents. For example, the word “summer” should be understood as two syllables, and the accent situation is different – the re-read “sum” and the non-re-read “mer”.
When Deep-speare writes sonnets, the language model generates a candidate line, and the rhythm model picks out the rhythmic lines of the poem, and repeats the process.
And, of course, there’s a rhyme model. The model focuses only on the last word of each line of the poem, as much as possible to achieve the word rhyme. For example, “day” and “may”, “temperate” and “date”.
When examining the poetry output, the team found that deep-speare generated poetry phrases that did not overlap much with the training data. In other words, instead of memorizing the training data and copying it directly, it created original poetry. However, this does not explain the literary quality of poetry.
To this end, the research team found two critical committees, allowing the judges to distinguish between humans and machines to create 14 lines of poetry.
The first were crowdsourcing workers hired by Amazon Mechanical Turk, who only spoke basic English but had no expertise in poetry. The end result is that workers guess human poetry and machine poetry with 50% accuracy. However, this figure may be false because workers may search the internet for excerpts of poetry, human poetry will have feedback on search results, and machine poetry will not appear.
The second judge is Adam Hammond, an assistant professor of literature at the University of Toronto. This time, the way to judge is to rate the rhythm, rhythm, readability and emotional impact of the son-and-thin lying son-in-law. As a result, Adam Hammond rates Deep-speare’s rhythm and rhyme severance, even beyond humans (because poets often deliberately don’t follow the rules for some effect);
Currently, the research team is working to improve the performance of AI poets in terms of readability and emotional impact.
Human poets, on the other hand, don’t sit at a table and think, “Well, what should my first word be?” Then, after making this difficult decision, rack your brains for a second word. Instead, the poet has a desired subject in his mind and then looks for words to express the idea.
The research team has taken a step in that direction by giving Deep-speare the ability to generate poems based on specific themes, such as love or loss. Moreover, sticking to a topic can increase the consistency of four lines of poetry, and the words that the model can choose will be limited by the subject.
There is no doubt that this is an ambitious project.