GARTNER, INC. :“By 2018, 20% of business content will be authored by machines.” We: “So is poetry!”
What it does
Given any topic supplied by the user, this skill composes a new, custom poem about the topic.
How it was built
Our backend system is based on the paper “Generating Topical Poetry; Marjan Ghazvininejad, Xing Shi, Yejin Choi, and Kevin Knight; EMNLP 2016”. We implement an API that accepts any topic phrase and returns a new poem composed by computer in real time.
Challenges we ran into
- Designing the utterance user interface for Alexa.
- Accelerating our backend so that it can generate a four-line poem within 1 seconds.
- Tune the hyper parameters with human in the loop.
- Coming up with a skill name which is easily recognizable for Alexa.
Accomplishments that we are proud of
- Speed up the whole pipeline using GPU (30s to 1s for generating a 4-line poetry - alexa motivated this speedup)
- Find a better hyper parameter by collecting a large amount of feedback from Amazon Mechanical Turk.
What we learned
- Build a closed pipeline using : Amazon Web Services + Amazon Mechanical Turk + Amazon Alexa Skill.
- CUDA GPU acceleration technicals.
What's next for Custom Poetry
- Maybe short stories on topics
- Maybe poems for other languages like Spanish.