A quick look around the SoDash office will confirm that fashion and beauty, have never been our strongest suit. A Beautiful Mind, maybe. A Beautiful Man? Well. We’re working on it. However, what we are passionate about, is creating beauty in technology. Now, let us question – what is beauty? This is a question that philosophy has failed to solve, so it’s very unlikely that a tech blog is going to advance this forward significantly. The Hippias Major, written by Plato, documents a series of discussions between Socrates and Hippias, to define “kalon” – a word that approximates, but doesn’t quite equal, “beautiful” in English. As is the case in almost all Greek philosophy, particularly that of Socrates, the conversation ends without resolution. However, one of these definitions revolves around the concept of being appropriate. In essence, something is beautiful if it completely satisfies the description of what an item should be. For example, a cup of coffee could be classed as beautiful, if it possessed all the enjoyable elements of the drink to a particular individual – some expect certain smells, reactions, tastes, sounds, and other pleasures brought by it. It is only classed as beauty, when it achieves exactly what we desire it to.
That wasn’t bad, was it? Indeed, it almost sounded like we know what we’re talking about (spoiler alert: we don’t). However, in a fairly recent experiment in our downtime, we set ourselves a challenge – what if robots, could write poetry? Would they exceed expectations, and find that level of beauty – in this case, the beauty of hand-crafted poetry tailored to the individual. Exciting, don’t you think? Well, we thought it was. But why poetry? At SoDash, we’re passionate about understanding human interaction, and moulding our systems to accommodate and manage it in the best way possible. Voltaire once described poetry as “the music of the soul”. Wordsworth, more eloquently and accurately, describes it as “the spontaneous overflow of powerful feelings.” Sentiment analysis and emotion management is what makes us tick, so naturally, poetry was the best test.
We therefore set out to create our Haiku AI. What follows is going to be a little alienating to the casual reader, but for those interested to know how we did it, here’s a quick rundown. If not, skip to the next paragraph – it’s not often we’ll actively choose to be ignored, but in this instance, it feels fair. The Haiku corpus was developed, through the use of a POS tagger, and the development of grammatical skeleton fragments. Through an n-gram model and the creation of topic vectors, the general text corpus for the AI to be based around was created, which combined with the skeleton fragments made a full Haiku template. We then proceeded to assign syllable counts to slots, and filled it in using n-gram and topic-related scores – resulting in, your tailored Haiku. Onward, to testing…
So, how did we do. The answer is, not that bad. We ran a survey, and generated a Haiku for 4 topics, and for each topic 4 poems: one competition winning human haiku, one “random” human haiku (picked from Reddit’s haiku thread), one randomly sampled AI haiku, and one where we generated a set of AI haiku and picked the best. We then proceeded to a Turing test, to see if people could distinguish between man and machine. Overall, people could distinguish whether the author was human or computer. However, 31% considered the AI poems to have been human written – compared with 67% for actual human poems. Indeed, this rose to 43% for the quality-filtered AI poems, and of the 8 AI poems in the evaluation, one received a majority verdict that it was human-written. Reassuringly, of the good AI poems, 84% were considered on-topic, compared to 78% of good human poems – so while we lacked the human beauty, we out-performed human focus.
Where does this lead us? The haiku is a blend of two separate concepts, fused together. We hope to see this moving towards concept blending, and managing multiple concepts to find derived meanings for each individual component of speech. We also hope that this moves towards collaborative AI, enabling AI-to-AI, and AI-Human collaboration. It also helped inspire our work on language processing and automation, SoGrow. For those interested to learn more, and review the project yourself – you can access it by clicking here.