Extracting Knowledge from Natural Language Documents to Achieve Competitive Advantage
Tsukahara’s research on document intelligence and DX innovation
With the proliferation of generative AI, large language models (LLMs) have rapidly become a mainstream research field. At ITLAB, our research on natural language processing and LLMs primarily focuses on improving existing LLM models and their practical applications. My current work involves applying LLMs to enhance the efficiency and quality of document creation—such as specifications and design documents—across various departments at DENSO as part of the company’s digital transformation (DX) initiatives.
As a component manufacturer, DENSO develops and produces products based on requirement specifications provided by various automobile manufacturers. When specification documents are of poor quality, development rework occurs, leading to increased costs and reduced product quality. Furthermore, specification writing styles differ by manufacturer and product category. Currently, these documents are checked manually, but if LLMs could inspect and suggest corrections without human intervention, we could eliminate human oversights and inconsistencies caused by varying skill levels.
This research originated from a consultation about effectively utilizing DENSO’s accumulated internal documents. Because design and requirement documents written in natural language are unstructured, it has been difficult to search and leverage them effectively. Initially, we explored retrieval techniques using statistical language processing to surface valuable internal knowledge. If we can use machine learning and LLMs to structure the knowledge embedded in documents, we could eventually aim for a system where people write requirements in bullet points and the system automatically generates requirement, design, and test specifications. This would improve document quality, increase development efficiency, and ultimately enhance product quality—yielding a competitive advantage over other companies.
LLMs in Recommendation and Question Answering Systems
Tachioka’s work toward smarter industrial applications
My LLM-related research in natural language processing focuses on two main areas: recommendation and question answering. For recommendation, much like how music streaming services suggest songs based on listening habits, we aim to leverage the extensive knowledge within LLMs to improve recommendation accuracy. The second area, question answering, explores enabling LLMs to correct errors in reference materials using their knowledge so that they can provide accurate answers even when the source documents contain mistakes.
In recommendation applications, suggesting “what to do next” within a system can enhance user experience (UX). For instance, in a manufacturing process, an LLM can recommend the next task using its knowledge, preventing tasks from being missed or overlooked. Question answering, meanwhile, can support summarizing long documents—such as specifications—by topic, to prevent errors from being introduced. Since LLMs can converse smoothly yet are prone to hallucinations—producing responses that deviate from facts—researching fine-tuning techniques to mitigate these issues is also part of our work.
Because LLMs are trained primarily on web data, they struggle to grasp physical phenomena in the real world and the common sense we possess in daily life. They also lack internal corporate knowledge and domain-specific expertise not published online. To address this, we are working on Retrieval-Augmented Generation (RAG), which supplements an LLM with external information it does not originally possess.
Lowering the Barrier to NLP and Expanding Integration Possibilities
How LLMs reshape Japanese language processing
Before LLMs, Japanese natural language processing required “morphological analysis” to segment sentences into words, followed by “dependency parsing” to infer structure and meaning. With LLMs, we can now accurately infer what a text is about simply by providing it, without such preprocessing. This has significantly lowered the barrier to entry for NLP and made integration with other systems much easier.
However, the black-box nature of LLMs presents challenges: when an error occurs, it’s difficult to identify where and why. Even obvious mistakes can appear correct at first glance, making diagnosis hard. As a workaround, we currently rely on prompt engineering, where humans work hard to prevent errors through careful instruction. How prompts are designed and phrased has become a key factor in successful applications. While we continue exploring prompt design through trial and error, such approaches lack generalizable algorithms or methods, which makes it difficult to call it research. If we could develop LLM agents capable of autonomously generating and processing prompts, that might become a research subject.
Toward Conversational AI and Robot Control
From dialogue systems to real-time human-machine interaction
My entry into NLP research started with dialogue systems. I worked on conversational systems designed to prevent drowsy driving through casual conversation, as well as providing entertainment content such as local sightseeing or gourmet information. In the past, conversations often broke down or the responses were off-target, but with LLMs, natural and sustained dialogue is now achievable.
Research applying LLMs to robot control is also attracting attention. If we imagine an AI that converses continuously with the driver while operating a vehicle—like KITT from *Knight Rider*—that’s the future we’re approaching. Many researchers in automotive UI/UX are likely aspiring toward that vision.
Future Research: Breaking Through the Knowledge Limits of LLMs
Bridging physics, language, and intelligence
What I find fascinating about natural language processing research is that, unlike fields such as signal processing or image recognition where results are purely numerical, improvements in NLP are often *intuitively apparent*—you can tell immediately when something “sounds clearly better than before.”
I take the opposite view. I find quantitative research easier to interpret, while language is complex because its meaning varies by individual perception. Although quantifying linguistic phenomena is difficult, that very challenge makes it intriguing. The same statement can be perceived positively or negatively depending on context. There’s no single right answer, and when you dig into why, perspectives from folklore and cultural anthropology emerge. That diversity of interpretation makes the field especially captivating.
LLMs’ robustness to diversity—being able to fully leverage performance by skillfully managing variations in both input and output—was unimaginable just a few years ago. In earlier dialogue research, we manually created conversational variations, but beyond a certain point, it became ineffective. With LLMs, however, performance continues to scale as parameters increase.
Conversely, since LLMs have already trained on nearly all publicly available web data, their knowledge expansion has plateaued. To overcome this limitation, future research will likely focus on self-training, where LLMs generate their own training data, and on AI alignment, which fine-tunes broadly trained models to align with human intent and values.
While applying LLMs is fascinating, I’m also eager to investigate what’s happening inside LLMs themselves. Since statistical machine learning shares the same framework as statistical mechanics, examining them using physics knowledge could yield insights that advance both theory and practical applications.
Join Our Research at DENSO ITLAB
Where engineering curiosity
meets real-world impact
