AI Researchers and Haiku Poets Why Do People Create Haiku?

Web Technology Digital Transformation Technology Artificial Intelligence Technology Natural Language Processing Technology Semantic Web Technology Deep Learning Technology Online Learning & Reinforcement Learning Technology Chatbot and Q&A Technology User Interface Technology Knowledge Information Processing Technology Reasoning Technology Zen and AI Navigation of this blog
AI Researchers and Haiku Poets Why Do People Create Haiku?

From AI Researchers and Haiku Poets From Why Do People Create Haiku?

AI研究者と俳句

This book will be written by Professor Hidenori Kawamura, who conducts AI research at Hokkaido University. Professor Kawamura has also created a system of AI willow poems, which can be found on Twitter at the Kawamura Lab (Harmony Lab). For example, there is a haiku about world peace, “Laughing together in prayer for peace among mankind,” and another about the piano, “My wife hasn’t played the piano for a long time.

Professor Kawamura has created an artificial intelligence “Issa-kun” that uses deep learning to learn haiku written by a haiku poet (Kai Otsuka), and has asked Mr. Otsuka to select the best of the haiku and then discusses the content with the source of the teacher data (the haiku poet). The purpose of this book is to consider the following issues in artificial intelligence research through these tasks.

  • How do people conduct the world’s acknowledgement of the world?
  • How do we connect information and concepts from the real world to symbols?
  • What are emotions?
  • What does it mean for a communication to be established?
  • What is love, the basis of human relationships?
  • What is the meaning of physicality for intelligence?
  • What is the future in which humans and artificial intelligence coexist?

In the discussion with the haiku poet (Mr. Otsuka), when he looked at the haiku generated by “AI Issa-kun” based on his own haiku, he felt as if he were in the haiku, which was clearly a strange sensation different from when he selects haiku from others. In the discussion of the significance of seasonal words in haiku, we consider “common knowledge,” which is a prerequisite for human communication, and in the discussion of communication through the language of haiku, we consider how computers handle analog information in terms of encoding/decoding. In his discussion of communication through the language of haiku, he contrasts it with the encoding/decoding of analog information on a computer, and deepens his examination of what haiku poets enjoy in their haiku.

And finally, there is a discussion on the topic of the haiku version of the Turing Test, in which an artificial intelligence is invited to participate in a haiku gathering, and when the haiku is submitted and selected by people, the haiku is tested to see if it can be detected as an artificial intelligence, the significance of creating haiku with artificial intelligence, and how to handle love in artificial intelligence and haiku.

Professor Kawamura states that through these conversations with Mr. Otsuka, both the question “Why do people write haiku?” and the interest in “How can we create haiku with artificial intelligence?” emerge from the workings of “intelligence” itself.

The following describes the details of this book and the mechanism of haiku generation by “AI Issa-kun,” which was described as an appendix after the book.

Chapter 1: Codes and Chronicles – Unraveling “Human Activity

For AI researchers, for haiku poets
Mysterious experience of selecting haiku
The Fundamental Why
Season words as Tips
Sabaun Not a Thing to Tell People Akiya Kato
The Legacy of the Chronicle
Till the field is blue with dreams of sleeping well Kabutota Kaneko
AI has no common sense
Love’s agility and AI
Interpretation changes with the times
Encoding author, decoding reader
A single phrase transcends time and space
Composition is digital, reading is analog
The author and the subject of the composition

Chapter 2: Symbols and Meanings – What are the hurdles?

Haiku made without knowing the symbols
Can you visualize “sad”?
Can you visualize “particle”?
Even if you can create an inexhaustible number of haiku, you cannot select the best ones.
Quantify excellent haiku by “features

Chapter 3: Teacher data and deviation – Let’s appreciate “AI Issa-kun”‘s haiku.

Finding Honka-tori and Parody
Approach to sexual love motifs
Strongly felt male perspective
Bias of teacher data
Is “word play” and “rhyming” possible?
Good at proper nouns
Spoken language appears unexpectedly
Refrains perceived as useless
Excess words” that are played off
Small variation in evaluation
Corrections are done by AI, but polishing is done by humans

Chapter 4: The Turing Test and Haiku Meetings – Will the day come when we can say we have “composed” a haiku?

Is Haiku about “understanding others” or “understanding oneself?
The day when AI appears in online haiku gatherings
Is it “compose”, “write”, or “generate”?
Where does the aversion to AI haiku come from?
Does it make sense for AI to compose haiku?
The comfort of words circulating like handprints
Possibility of AI ending haiku
Foregrounding context
The Dynamic Equilibrium of Sequences
The currency of seasonal words

Chapter 5: The Unconscious and Emotion – The Fundamental Question of “What is Intelligence?

The growth process of others
What we see when we raise the level of abstraction
Are we “writing” or “being written”?
What is interesting and how is simple
I don’t know, but it’s interesting
AI, which cannot fall in love, talks about love
Philosophical question with no answer

Appendix: How “AI Issa-kun” Haiku Generation Works

How “AI Issa-kun” Generates Haiku

There are two main processes of haiku generation by AI Issa-kun. First, AI Issa-kun selects words that compose a haiku one by one and connects them in order to generate a sequence of words that could be used to form a haiku. Next, we estimate whether the words in the sequence satisfy the conditions for a seasonal haiku and whether they make sense, and select only those that satisfy the conditions.

To do this, Issa-kun uses a “language model” based on deep learning.

Unlike most language models, the language model used in Issa-kun is based on Japanese literary works as teacher data. The limited number of haiku alone is not enough to provide enough data to teach the relationship between words, so the model is first trained on data that is thought to have something in common with haiku, with an emphasis on describing scenes. For example, three sentences with five, seven, and five syllables are included in the teacher data.

Next, in order to learn the unique characteristics of haiku, such as the restriction of the number of syllables and seasonal words, additional data of haiku written by others are learned and fine-tuned. The procedure for generating haiku is as follows.

    • Select and arrange the words in order from the top of the list.

When Issa-kun generates haiku, it first uses the document generation model learned to generate haiku to create a sequence of words that looks like haiku. The probability of each word being represented is calculated using a neural network, and the computer generates sentences by randomly selecting and arranging the words in order from the top, using dice that produce eyes according to that probability.

Ordinary dice are designed so that each die has an equal probability of coming out, but the dice (random numbers) rolled here are differentiated according to the probability calculated by the sentence generation model, with some words more likely to come out and others less likely to come out. Words that the sentence generation model estimates have a high probability of coming first are more likely to be selected, while words with a low probability are less likely to be selected.

The model calculates and determines the probability of the second and subsequent words in the same way, but unlike the calculation for the first word, the probability varies depending on various conditions, such as what words precede it, how many syllables it has, and whether or not a seasonal word has already appeared. Therefore, the document generation model takes all of the word sequences selected so far and calculates the probability of the next occurrence of a word according to the sled.

    • Removing Haiku that do not match the conditions (1) Are they too similar to the teacher data?

It is difficult to say that the sentence generation model is able to learn perfectly from haiku written by people, and the sentences produced by Issa-kun’s sentence generation model often do not include seasonal words, do not seem to have 17 syllables, or produce sentences that do not make sense. Conversely, the sentence generation model may generate exactly the same haiku as the one it learned.

In order to eliminate such incomplete haiku as much as possible and to produce only high-quality haiku in the end, Issa-kun evaluates the generated haiku and removes those that do not meet the conditions.

First, the generated haiku are compared one by one with the haiku used as teacher data, and those that are too similar to the teacher data are excluded from the Issa-kun generated results. Haiku generated by Issa-kun are excluded not only when they are exactly the same as those written by others, but also when they are clearly similar, such as when only one letter is changed for a girl.

Therefore, we calculate the edit distance between the haiku and the teacher data, and exclude haiku that are less than a certain distance. The edit distance is defined as the number of times a sentence can be reconstructed from one to the other by adding, deleting, or changing characters at least a few times when the two sentences are compared.

    • Remove haiku that do not meet the requirements (2) Do they meet the requirements for a definite form for a certain season?

Next, remove the haiku that do not meet the conditions of a fixed form for a given season. To determine whether a haiku is a seasonal form, it is necessary to examine the number of sounds in the words that make up the haiku and whether the words in the haiku are seasonal or cut-off characters. The National Institute for Japanese Language and Linguistics (NINJAL) has published a dictionary that classifies the various words that appear in the Japanese language, and this dictionary is used to divide the words in the generated haiku. The smallest unit of the divided words is called a morpheme in the terminology of Japanese linguistics, and the technique of automatically dividing sentences into morphemes is called morphological analysis. The dictionary assigns a reading kana to each morpheme, from which the number of syllables can be counted.

If an unknown morpheme is found among the morphemes of a generated haiku that is not in the dictionary, it is assumed that the sentence generation model has inadvertently generated a word that is not in the original haiku, and the haiku is excluded. If the morphological analysis results indicate that the number of morphemes in the generated haiku does not total 17 sounds, or that the haiku contains morphemes that span between five, seven, and five sounds, the haiku is removed in the same way.

Issa-kun also has a pre-registered dictionary of the seasonal words and clipped words used in haiku. The number of seasonal words and clipped words in the generated haiku are counted by comparing the morphemes in the haiku with this dictionary. Haiku that contain no seasonal words, haiku that contain two or more seasonal words, and haiku in which a sekiji appears two or more times are excluded from the general type of seasonal form haiku, since they do not meet the conditions.

    • Haiku that do not meet the condition are eliminated. (3) Are the haiku meaningful?

In addition, we estimate whether the haiku generated by Issa-kun are meaningful or not. Using the generative model used to generate the haiku, we can estimate the probability that the sequence of words will be broken throughout the generated haiku, and the higher the probability, the more likely it is that the haiku is meaningful.

While the haiku that the sentence generation model judges to be good do not necessarily make sense, there are many haiku that it judges to be bad that clearly do not make sense, indicating that these evaluations are working to some extent.

  Challenges to go to the next level

One of the remaining challenges for “AI Issa-kun” in the future is how to express the “message” before the haiku and how to convert it into words. Currently, it is generating haiku from nothing, which means that it is avoiding the problems of encoding and decoding, i.e., generation and interpretation. In order for “AI Issa-kun” to go to the next level, it is necessary to solve how to handle the selection and criticism of haiku.

Another major issue is the need to consider how to handle multimodal information and how to link information in different modes. We can only have shared knowledge with others when information in various modes comes and goes, such as when we put into words the sensations we see with our eyes or feel on our skin, or conversely, when we imagine sensations from words.

Introduction to Haiku Reading

Sound and Rhythm as the Core of Haiku
The Mechanism of “break
What “kiri-ji” suggests
Why “kigo” is so important
The battle against form

Key words to get a general idea of the contents of this book

コメント

Exit mobile version
タイトルとURLをコピーしました