The potential of AI in technical communication
Artificial Intelligence (AI) refers to the simulation of human intelligence through software. In recent times, AI has gained a lot of attention as it offers the potential to automate manual tasks. However, what real benefits and potentials does AI actually bring to technical communication?
End of 2022, OpenAI released the groundbreaking chatbot ChatGPT. With its ability to generate human-like text and answer complex questions, this chatbot has raised high expectations regarding intelligent delivery of accurate information. However, the innovative power of this technological advancement also raises the question of whether we need to discard proven approaches and methods in technical communication, from metadata to semantic information models and knowledge graphs.
Generative AI and Large Language Models
To explain the potential applications for technical communication, we would like to introduce the terms “Generative AI” and “Large Language Models.” Generative AI refers to a form of artificial intelligence capable of creating new data or content that is similar to the data present in its training dataset. Complex models and algorithms are used to generate texts, images, or videos that could have been produced by humans.
ChatGPT is an example of a Generative AI. The chatbot, based on a generalized pre-trained transformer, provides astonishing real-time responses to various user prompts. This capability is made possible by Large Language Models (LLM), which have been trained on massive datasets to process and analyze language. Words are related to each other through statistical calculations and represented as vectors (embeddings). The calculated proximity between two words correlates with their semantic closeness, although the language models do not possess true language comprehension or understanding of the world in the conventional sense.
Based solely on a language model, we believe it is not advisable to reliably provide customers with precise and often safety-critical technical information, such as repair instructions. However, with the hype surrounding these technologies, where do we see potential applications and benefits? We have identified and tested the following scenarios in practice.
Applications in technical communication.
A “classic” task of technical communication is the provision of information (documents and topics). Unfortunately, these, especially when dealing with legacy data, generally have no or few metadata, making targeted searching in the delivery systems (e.g., through facets) not possible. Large Language Models (LLMs) offer a very good opportunity here to analyze and classify information stocks. The results of the classification become particularly powerful when LLMs are supplemented with factual domain knowledge, for example, through knowledge graphs and/or trained for a specific task (keyword: fine-tuning with ground-truth data).
Frequently, knowledge is embedded in various documents within a company. Large Language Models have proven to be very useful in the extraction of knowledge from documents, such as technical specifications of products. Why? Because based on the user prompt, they are capable of precisely identifying and extracting the relevant information. The extracted information can then be fed into a knowledge graph, for example. The process of populating a knowledge base can thus be automated for both structured and unstructured information. Additionally, these services improve when the implicit knowledge of the Large Language Models is complemented with explicit, structured knowledge, for instance, from a knowledge graph.
Image: Information Classification and Information Extraction with Large Language Models and Knowledge Graphs
Dialog-driven search and troubleshooting.
In various projects, we use knowledge graphs to connect verified information from different information silos. Through an application, we provide this information to users. We will continue to follow this approach in the future because only the content approved by relevant departments is reliable. At the same time, finding the required information is not always easy. We see great potential in using Large Language Models (LLMs) to guide users to the right information in question-and-answer scenarios without generating the answer itself. This is particularly relevant for activities like troubleshooting or diagnostics.
This approach is still in its early stages: Based on existing content, for example, in a private instance, ChatGPT can generate new content. Considering product liability and safety, this approach may seem risky. However, what if ChatGPT could assist technical writers in the future by providing a draft, let’s say for a newly created topic? The task of technical writers would then be to elaborate, verify, and approve this draft. This ensures that only quality-assured content is provided. The content creation time can be significantly reduced by using ChatGPT.
Preserving the Proven - Metadata, Semantic Information Models, and Knowledge Graphs
We keep a cool head amidst all these hypes. We bring clarity to the technology jungle and guide you on the path ahead. We hold on to what has proven to work while complementing it with new technologies like LLMs and ChatGPT where it makes sense. This way, you can optimize your processes and provide your customers with the right information tailored to their needs.
Our approach is built on various pillars, depending on your current situation and the strategy you pursue. Metadata, semantic information models, and knowledge graphs are not outdated concepts. They help us, and in turn, help you intelligently deliver information. Moreover, they also assist language models in improving their performance because, as the saying goes, the better the input, the better the output.
By the way, it’s worth mentioning that recently, American researchers discovered that GPT’s answers have degraded over time.
Subscribe to the free newsletter from PANTOPIX. We will gladly keep you informed regularly about new articles.
Combining different data sources in a knowledge database and the semantic representation of the information contained therein can make technical communication much easier. Building a knowledge base using semantic knowledge graphs offers numerous advantages, including the important possibility of continuously expanding the knowledge graph. One method of expanding knowledge is the use of knowledge graph embeddings.
The world of software development is constantly changing, and with it the way we document software. One innovative method that has gained popularity in recent years is “Docs-as-Code” or documentation as code. In this article, we will explore this exciting development and take a closer look at its application in technical documentation.
Prof. Dr. Martin Ley
Partner and Senior Consultant