If we wish to generate haiku, we must first understand what haiku are. Since its exposure to western audiences, its popularity as a poetic form has veritably exploded. As such, there are many online resources discussing the intricacy of the form. This is intended to be a general overview, as any attempt to do otherwise would do the form a disservice.
Haiku are a form of Japanese poetry. In their original form, they were the first three lines of a waka, also known as a tanka or uta. Waka are a five-line, 5-7-5-7-7 poem, that are often strung together into renga, which are a linked verse dialog between multiple poets.
Of these long poems, composed of many waka, the first three lines are called a hokku. The hokku are the most recognizable and impressionable verses of renga, and are analgous to the first verse and chorus of a song. Renga were composed in meetings of several poets, and it was common for poets to come prepared with pre-composed hokku.
These hokku are what is now referred to as haiku.
Upon sharing my work in haiku generation, the very first question that arises often has to do with the 5-7-5 syllabic structure. The common perception of haiku, is that it is a poem, composed of three lines of 5, 7, and 5 syllables each, with no further constraints. This is only partially true.
Haiku maintain a thematic structure as well.
It is difficult, even impossible, to maintain the original syllabic structure in translation from Japanese to English. Thus many translators have striven to match the style, theme, and mood of a Japanese haiku, and while maintaining its three-line structure, they abandon the strict syllabic requirement. Further, note that, as with any style of art, there was experimentation with a freer verse style, both in content and structure, long before the haiku grew to popularity in western culture.
Joan Giroux defines
A haiku is a 17-syllable poem arranged in three lines of 5, 7, and 5 syllable, having some reference to the season and expression the poet's union with nature.
in The Haiku Form.
Regardless of its definition, the dataset I have collected does not exhibit strict syllabic structure, but does seem to exhibit the strong seasonal and natural aspects that Giroux describes. Because a neural network model is only as good as its training dataset, I do not expect the generated haiku to follow the 5-7-5 syllabic structure, but I do expect them to exhibit strong natural themes.
Further, what interests me about this project has little to do with the particular poetic form under consideration, but the ability for a machine learning model to generate art, and in that respect, I believe my efforts have been categorically successful.