How Alibaba’s Tmall Genie Achieves Human‑AI Co‑Creation in Poetry

After AlphaGo sparked interest in AI imitation learning, concerns grew about job displacement, prompting Alibaba’s AI Lab to showcase human‑AI co‑creation through the "Tmall Genie writes poetry" service, which lets users generate acrostic love poems on a smart speaker.

The AI creation team, part of A.I.Labs' cognition lab, explores AI creative technologies, deep learning, natural language understanding, and intelligent dialogue applied to smart speakers.

Why focus on AI poetry for co‑creation? The team believes that combining intelligence with emotion via human‑AI collaboration raises the artistic quality of works, allowing machines to better understand emotional expression.

How is the co‑creation implemented? The system offers word‑level candidate sets for each character and uses beam search to fix parts of a line while generating subsequent text, providing recommendations rather than letting users edit the whole poem alone.

To avoid quality loss when modifying a single line, the model keeps placeholders for each line, generates multiple candidate expressions, and selects the best combination using a unified evaluation model.

How is the evaluation model built? The overall poem quality is represented by the poem’s probability, decomposed into bigram bag‑of‑words embeddings fed into a fully‑connected network. Each edge (pair of consecutive lines) receives a score; the poem score is the product of all edge scores. Pruning removes candidates that violate rhyme or tonal rules, and a max‑K path search finds the highest‑scoring quatrain.

How does the system ensure context continuity? It uses an RNN (LSTM) encoder to embed the semantics of the previous line and a decoder RNN language model to generate the next line, with an attention mechanism selecting the core meaning from the encoder output.

The team believes that collaborative AI‑generated poetry offers higher quality than previous standalone generators, encouraging users to modify lines and participate in the creative process.

What are the future prospects? Language understanding and generation can extend to broader creative scenarios, turning AI into a collaborative partner that opens new possibilities for literary and artistic creation, much like AlphaGo opened new strategies for chess.

Ultimately, AI should assist rather than replace humans, and human‑AI co‑creation may be the key to unlocking AI’s full potential.