The True Nature of the "Response Discomfort" Felt in Japanese LLMs — Reasons Why Responses Don't Align and Approaches to Improvement

Table of Contents

• 1. Introduction
• 2. Why Does the "Response Discomfort" in LLMs Arise?
• 3. The Unique Challenges of Japanese and Company-Specific Contexts
• 4. Model Limitations or Design Issues?
• 5. Can a Single Company Improve Response Quality?
• 6. The Management Challenge of "Response Quality Design"
• 7. Conclusion
• 8. Human Science Teacher Data Creation, LLM RAG Data Structuring Outsourcing Service

1. Introduction

As the business use of LLMs, including ChatGPT, progresses, there are increasing opportunities to hear voices saying, "It's convenient, but somehow it doesn't quite fit." The sentences are well-structured. The honorific language is natural, and the explanations seem plausible at first glance. Even so, when trying to use it in actual work, a subtle sense of discomfort remains. The answer you wanted is somehow off.

This sense of discomfort may not be explained solely by the skillfulness of the prompt. In this article, we will organize how to understand the issue of Japanese LLM responses not aligning well, and what can be done to address such problems.

2. Why Does the "Response Discomfort" in LLMs Arise?

When you feel a sense of discomfort with an LLM's response, in many cases, the output is not actually incorrect. It is also natural Japanese. So why does it still feel off? This can be attributed to the fact that the criteria for evaluating the quality of the output cannot be measured simply by right or wrong.

In business, there are as many various assumptions and tacit knowledge as there are companies. These include, for example, relationships with customers, the history of past transactions, and industry-specific customs. It could even be called culture. When humans interact, conversations proceed with mutual understanding even without fully verbalizing such culture. However, the scope of what LLMs have learned does not include these company-specific cultures and is based on general information.

As a result, it is possible that the response generated is generally correct but not suitable for the company’s culture. This “mismatch in responses due to not taking culture into account” can be said to be one of the causes of the sense of discomfort in LLM responses.

3. The Unique Challenges of Japanese and Company-Specific Contexts

Japanese is a language where subjects and premises are often omitted, and it has a high dependency on context. Expressions such as "We will consider it" or "We will think positively" can have significantly different meanings depending on the relationship and situation.

Furthermore, each company may have its own unique standards. Even the phrase "respond promptly" can mean same-day response for some companies, while others may consider a response within three business days as acceptable.

On the other hand, LLMs learn from average data. Therefore, the unique standards and culture of a specific company are not automatically reflected. It is not hard to imagine that when the difficulty of the Japanese language overlaps with company-specific culture, the sense of discomfort becomes more apparent.

Reference blog: The Role of RLHF in Domestic LLMs — Where Does "Human Judgment" That Determines the Quality of Japanese LLMs Come Into Play?

4. Model Limitations or Design Issues?

When a sense of discomfort arises, the discussion tends to turn to whether the performance of the introduced LLM is insufficient. Naturally, this leads to the thought that switching to a larger-scale model might improve the situation, or that adopting a more accurate foundational model might resolve it.

However, if the performance of the LLM improves, will it automatically understand the company's culture and decision-making criteria? The reality may not be that simple.

Rather, the question should be not about the model's capabilities, but about how the response quality is designed. Here, design does not mean delving into the model's structure itself, but refers to defining the evaluation criteria for what constitutes a "good response" for the company.

For example, even in the way customer support is written, some companies present conclusions succinctly, while others carefully explain the background. The same applies to the style of internal documents and the approach to new proposals; the criteria for what is evaluated and what is considered good vary from organization to organization.

If these evaluation criteria are not clearly defined, no matter how high-performance the model you introduce is, you will not be able to escape the feeling of "something is off." The root cause of this discomfort may lie not in the limitations of the model, but rather in the lack of design for such response quality.

5. Can a Single Company Improve Response Quality?

So, is it possible for a single company to improve the response quality of an LLM?

To conclude, retraining the foundational model is not realistic, but improving response quality in your own business operations is definitely possible.

The starting point, as mentioned earlier, is response design based on your company’s own standards. First, it is necessary to clearly define what constitutes a "good response" for your company. Define the criteria by which you will evaluate the LLM’s responses and verify them using actual business data. Evaluate the LLM’s output using past customer inquiries and proposals to identify where it deviates from expectations.

On top of that, we refine the prompt design and incorporate a structure that references internal documents using methods like RAG as needed. Furthermore, by accumulating ideal answer examples and annotated data, it is also possible to correct tendencies through lightweight fine-tuning.

The important point is that improvement is not about making the model larger, but rather a process of teaching the LLM your company's decision criteria as data to correct response discrepancies.

Reference blog: The Key to Enhancing the Accuracy of Japanese LLMs is "Annotation Quality" ─ High-Quality Annotation Design and Operation Supported by Human Science

6. The Management Challenge of "Response Quality Design"

As the use of LLMs expands, response quality is no longer just a technical issue. It becomes a management challenge directly linked to brand, risk management, and customer experience.

What level of response is acceptable, and where will you not compromise? Define those standards, translate them into an evaluable form, and accumulate them as training data. Without this process, discomfort will not be resolved.

Behind the discomfort with LLM responses lies an urgent issue: the need for utilization that better fits practical business operations. What is needed to address this issue is not a debate over model selection, but rather the clarification of in-house standards and the design of response quality.

7. Conclusion

When you feel that the responses of a Japanese LLM do not align, it is not a sign of failure. Rather, it can be seen as an opportunity to clarify your company's unique value standards and strongly drive your business forward.

Instead of leaving the sense of discomfort unaddressed, verbalize where the misalignment occurs and clarify it as evaluation criteria. This accumulation is the path to adapting LLMs to practical use. Could this be the key to advancing the utilization of Japanese LLMs to the next stage?

Reference Blog: The Meaning and Future of Developing a Purely Domestic LLM

8. Human Science Teacher Data Creation, LLM RAG Data Structuring Outsourcing Service

Over 48 million pieces of training data created

At Human Science, we are involved in AI model development projects across various industries, starting with natural language processing, including medical support, automotive, IT, manufacturing, and construction. Through direct transactions with many companies, including GAFAM, we have provided over 48 million high-quality training data. We handle a wide range of training data creation, data labeling, and data structuring, from small-scale projects to long-term large projects with a team of 150 annotators, regardless of the industry.

Resource management without crowdsourcing

At Human Science, we do not use crowdsourcing. Instead, projects are handled by personnel who are contracted with us directly. Based on a solid understanding of each member's practical experience and their evaluations from previous projects, we form teams that can deliver maximum performance.   

Generative AI LLM Dataset Creation and Structuring, Also Supporting "Manual Creation and Maintenance Optimized for AI"

Since our founding, our main business and service has been manual creation, and currently, we also support the creation of documents optimized for AI recognition to facilitate the introduction of generative AI for corporate knowledge utilization. In sharing and utilizing corporate knowledge and documents using generative AI, current technology still cannot achieve 100% accuracy with tools alone. For customers who want to make the most of their past document assets, we also provide document data structuring. We offer optimal solutions leveraging our unique expertise, deeply familiar with various types of documents.

Secure room available on-site

Within our Shinjuku office at Human Science, we have secure rooms that meet ISMS standards. Therefore, we can guarantee security, even for projects that include highly confidential data. We consider the preservation of confidentiality to be extremely important for all projects. Even for remote projects, our information security management system has received high praise from clients, because not only do we implement hardware measures, we continuously provide security training to our personnel.

In-house Support

We provide staffing services for annotation-experienced personnel and project managers tailored to your tasks and situation. It is also possible to organize a team stationed at your site. Additionally, we support the training of your operators and project managers, assist in selecting tools suited to your circumstances, and help build optimal processes such as automation and work methods to improve quality and productivity. We are here to support your challenges related to annotation and data labeling.