How does the "Tmall Genie" Write Poetry for You? The Secret of Human-Machine Co-creation

During the Tanabata this year, the ""Tmall Genie wrote poems for you"" (hereinafter referred to as ""writing poems for you"") developed by the AI Creative Team of Alibaba Artificial Intelligence Laboratory (AILabs), through which you can use AI to give your beloved People write a Tibetan love poem, and they can also cooperate with human-computer co-creation.

Today, we have an exclusive interview with the head of the AI creative team. The team is responsible for exploring the application of artificial intelligence technologies such as AI intelligent creation, deep learning, natural language understanding, and intelligent dialogue on smart speakers in the cognitive laboratory of A.I.Labs.

How to realize man-machine co-creation in the creation of ancient poetry?

Q: Why do you want to cut in from the perspective of AI writing poetry, and do human-machine co-creation?

A: In the matter of literary creation, we believe that the most important thing is not intelligence. Human-computer co-creation can push the combination of ""intelligence"" and ""emotion"" to a higher level, making creation more ""mind"" and "" Emotion is more important. In the process of human-computer co-creation, the machine may also know more and more about the expression of ""emotion"". In order to prove the value created by man and machine, we have made an attempt to write poetry for the Tmall Genie.

Q: How do you create a poem for you?

A: The mode of man-machine co-creation is that the machine provides recommendations for people, rather than letting people complete the modifications alone. For example, the verse generation model can provide a candidate set for each word of each poem, and can also fix a part of the text by beam search to generate a post-order text, and the like.

However, the modification of a single sentence or a single word actually affects the quality of the whole poem. If only one sentence is modified and the subsequent verses cannot be recommended accordingly, the advantages of human-machine co-creation cannot be exerted. We have given a solution to this problem. Program:

In the process of generating quatrains, we maintain a different expression of each verse, and then choose matching sentences. We use a unified evaluation model to measure the quality of different combinations of good sentences, and choose the most quality as the output.

Q: How to build an evaluation model of human-machine co-creation?

The bigram word bag of the pre-order verse is scored after the embedded layer and then the entire sentence through the full-link network. This score can be regarded as the score for each side, and the score of the whole poem is the product of all the sides. The scoring model can be learned by training on the ancient poetry training set, and it is easy to speed up the caching of calculations during use.

The whole search process is a problem of finding the maximum K path, and the scoring of each path is done by the scoring model. During the search process, the pruning is used to remove the path of the rhyme and the inconsistency, and finally the best quatrains are obtained according to the ranking.

This process can start with any poem. When one of the poems is modified, you can rebuild the whole tree from it and then recommend the follow-up verse based on the results of the evaluation model.

Let the machine write poetry better

Q: In the matter of writing poetry, ""How do you write poetry for you?""

A: The biggest challenge from writing a poem to writing two poems is how to connect the context of two poems. That is to say, when writing the second poem, the machine needs to know what the first poem is saying, and it is inherited. .

The continuation of the context is to choose the appropriate Chinese characters to form a reasonable verse. It is necessary to solve the two difficulties of fluency and ambition. The former can be solved by the RNN model, while the latter needs to pass the meaning of the preceding sentence through the sequence-to-sequence model.

Sequence-to-sequence was first applied to the field of machine translation, and then the application scenario was extended to language generation tasks such as dialogue. The basic components of Sequence-to-sequence are two RNN networks: the LSTM Encoder network and the LSTM Decoder network.

The goal of the Encoder network is to represent the different elements of the first sentence as a real vector. It can be understood that the semantics contained in the first sentence are encoded into these vectors; the Decoder network is more like an RNN language model. However, in the selection of Chinese characters, consider the output of the Encoder network in addition to the first half of the sentence already written. When the output of the Encoder network is used, the central meaning of the preceding sentence is selected by Attention to achieve true satisfaction when generating the verse.

We believe that the quality of the whole poem should be improved by means of human-computer collaboration and co-creation, which is very different from the previous poetry program. The model of the machine is enriched, and the historical poetry is more elegant. In fact, it is not worthy of a sincere heart. So we encourage users to participate in the design and provide the ability to freely modify the verse.

Q: What is the application prospect of man-machine co-creation?

A: The understanding and generation of language can be applied to a wider range of creative scenarios. Putting a ""heart"" on the machine and creating it together with others will open up a new way for the peak of our artificial intelligence ""literary art creation"", just like the appearance of AlphaGo, giving professional chess players a more open-minded chess path to open new The same space.

Perhaps in the near future, the Tmall Genie can become a super input method. Instead of typing in each word you want to write, you can enter a synopsis or theme. The Tmall Wizard can generate a corresponding article and then modify it interactively. It's finished.

The value of AI is to help people, not substitutes, and human-machine co-creation and win-win may be the right way to open AI in the future.