Is Generative AI Weak Against 'No'? The Surprising Reason Japanese Prompts Fail

About this article

This article explains the reasons why the consecutive use of the Japanese particle "no" (の) in prompts hinders AI understanding, and provides practical methods for structuring them with parentheses and punctuation.

Why avoiding continuous "no" (の) drastically improves Generative AI response accuracy

Introduction: Why are Japanese prompts difficult?

"Tell me about the color of the leaves of the apple tree" and "Tell me about the 'leaf color' of the 'apple tree'."
Have you ever tried to see what difference these two prompts make in the AI's response?

In fact, this small difference greatly affects the AI's understanding accuracy.
If you feel that "Japanese prompts don't work well" or "I can't get the expected responses," the cause might lie in the linguistic characteristics of the Japanese language.

In this article, I will explain why simply being mindful of punctuation and parentheses can drastically improve your interaction with AI, using specific examples.

AI Confusion Caused by the Japanese Particle "No"

Humans vs. AI: Differences in Language Understanding

In the Case of Humans

• Understand meaning by integrating context, experience, and common sense.
• Can infer the "general meaning" even from ambiguous expressions.

In the Case of AI (Large Language Models (LLM))

• Calculate meaning based on statistical relationships between words.
• Determine "which words to focus on" using the Attention mechanism.
• Without a clear structure, accuracy decreases due to dispersion of attention.

Problems with Continuous "No"

The Japanese structure "A no B no C" is difficult for AI to process for the following reasons:

1. Ambiguity of Relationships: The modification relationships between A, B, and C are unclear.
2. Dispersion of Attention: It is difficult to judge which element should be emphasized.
3. Context Dependency: Interpretation relies heavily on the preceding and succeeding context.

Practice: Improving Accuracy through Structuring

Experimental Example: Comparing Apple Prompts

Let's compare how the AI's responses actually change with the following three prompts.

Pattern 1: Ambiguous Prompt

青森の農家のリンゴの収穫について教えて

AI Processing Status

• The four elements "Aomori," "Farmer," "Apple," and "Harvest" are lined up flatly.
• It is difficult to judge which element is the primary information.
• Tends to result in a generalized answer.

Pattern 2: Partial Structuring

青森の"農家のリンゴの収穫"について教えて

AI Processing Status

• Separates the location "Aomori" from the series of actions "Farmer's apple harvest."
• Recognized as a specific activity associated with a location.
• Answers containing more specific regional information can be expected.

Pattern 3: Full Structuring

"青森の農家"による"リンゴの収穫"について教えて

AI Processing Status

• Clearly separates "Aomori farmer" (subject) and "apple harvest" (action).
• Processing emphasizes the relationship between the subject and the action.
• More logical and structured answers can be expected.

Effective Use of Parentheses and Punctuation

1. Clarification through Grouping

❌ システムの設計の問題の解決方法
⭕ "システムの設計"における"問題の解決方法"

2. Explicit Prioritization

❌ 東京の会社の新しい製品の開発状況
⭕ "東京の会社"が手がける"新しい製品の開発状況"

3. Organizing Modification Relationships

❌ 最新の人工知能の技術の応用例
⭕ "最新の人工知能技術"の応用例

Why is This Method Effective?

Impact on Tokenization

Large language models process text by breaking it down into units called "tokens."
Parentheses and punctuation function as signals indicating "semantic chunks" during this tokenization process.

Concentration of Attention Weights

Structured prompts bring about the following effects in the AI's attention mechanism:

• Concentration of attention on relevant elements
• Reduction of unnecessary associations that act as noise
• Realization of more precise context understanding

Practical Techniques You Can Use Starting Today

1. "No" Chain Analysis

Identify parts in your prompts where the particle "no" (の) occurs three or more times consecutively and group them using parentheses.

2. Clarifying Subject and Predicate

Clearly indicate "who" is doing "what" to "whom" by using parentheses.

3. Step-by-Step Structuring

For long prompts, break them down into semantic units and organize them accordingly.

Conclusion: A Prompt is a Blueprint

A prompt is not just a simple question or command.
It is a "blueprint" for collaborating with AI.

While the flexibility of the Japanese language is convenient for humans, it can lead to processing difficulties for AI.
However, by adding visual cues such as punctuation and parentheses, you can significantly improve this challenge.

Action proposals for today:

• Check your existing prompts for consecutive "no" (の) instances.
• Try structuring them using parentheses.
• Observe and record changes in the responses.

Please share your prompt improvement experiences in the comments.
Let's all aim to improve the accuracy of Japanese prompts together!

---

References and Related Links

• Transformer Architecture and Attention Mechanisms
• Challenges and Prospects of Japanese Natural Language Processing
• Best Practices for Prompt Engineering

Next Preview
"Explaining 'Transformers' at the Heart of the AI Revolution in a Way Anyone Can Understand"

---

This article is a refined version of a past post from my blog, rewritten for Japanese readers.

https://imkohenauser.com/on-punctuation-and-parentheses-in-japanese-prompts/

If you have any other insights, please share them in the comments.