Haiku Generation From Narratological Perspective: A Circulation Between Haikus and Stories

Introduction

A haiku is a type of poem. This art form emerged in Japan and has been familiar to Japanese people for many years. A haiku poem is composed of 17 moras and 3 lines: the first line has 5 moras, the second 7, and the third 5. However, this standard definition is only a reference and does not define all types of haiku. Furthermore, many poets have attempted to describe haiku, with Matsuo Basho and Masaoka Shiki being some of the famous haiku poets. In Haikai Taiyo, Masaoka Shiki (1966) mentions that haiku is a literary form, pointing out that what makes it different from other types of literature is the number of moras, although haiku does not necessarily consist of 17 moras. In Haiku no Tukuriyo (“How to Make a Haiku”), Takahama Kyoshi (2009), an apprentice of Masaoka Shiki, comments on the need for seasonal words and cuts based on 17 moras. However, there is a view that haiku does not have a strict formal style. Ishihara (2012) thus asserts that a haiku style is defined on the basis of a poetic mindset (emotions, scenery, etc.).

As maintained by Masaoka (1966), haiku is a literary form, and it shows a high affinity for our research on generating stories. In this chapter, we propose a system for the following two functions and implement a prototype. The first function is to produce multiple haikus from a single story. The second function is to generate multiple stories from haiku.

The authors have approached the haiku production using two methods. One of them is grounded in the structural aspects of haiku, while the other are based on scrutinizing the transitions of the parts of speech in haiku.

Ito and Ogata (2019) consider that a haiku consists of parts of speech, and they created haiku using the transition patterns of parts of speech in each line. Specifically, this approach utilizes the transition patterns of nouns and verbs only (Ito, Igarashi, & Ogata, 2018). The target of analysis to acquire the transition pattern is 63 phrases published in Oku no Hosomichi (Matsuo, 1702; Imoto, Hisatomi, Hori, Yamashita, & Maruyama,2008). First, each phrase is subjected to a morphological analysis, and only the nouns and verbs of the obtained parts of speech are examined. The haiku is a fragmentary set of words and is an experimental effort based on the expectation that nouns and verbs make up the majority of it. In this paper, we focus only on the number of occurrences of nouns and verbs. If other parts of speech (e.g., adjectives) appear first, we begin counting from the first noun or verb that appears in the line. The second strategy for constructing a haiku uses a time series based on deep learning and long short-term memory (LSTM). The third strategy uses umeji, which harnesses the method of devising a story. The authors concentrate on a strategy of dividing the haiku into word and character units, grounded in the approach of deep learning employed by Ono and Ogata (2018, 2019). In each type of learning, the 63 phrases in Oku no Hosomichi were used to study the data. Igarashi harnessed a “chainer” for the framework of deep learning and produced a haiku by learning to construct the LSTM. The learned pattern is the appearance pattern of the division unit. The target haiku was given a start symbol and end symbol. The mechanism was taught 10,000 times under the above conditions. The authors tried deep learning for the appearance pattern of words in a haiku; the appearance pattern of characters in the haiku for the learning method is described. With these learning data, a haiku was formed by concatenating words or characters from the start symbol to the end symbol. The basic procedures are as follows:

• 1.

  Begin using the start symbol.

• 2.

  Select the element based on the learning data for the element that appears next to the start symbol.

• 3.

  If the end symbol is selected, the haiku is complete. Otherwise, go to Step 4.

• 4.

  Choose elements based on learning data for elements that appear next to the selected element. Return to Step 3.