Artificial Intelligence in Technical Documentation, Part 1
Artificial Intelligence in Technical Documentation (/software/) is already widespread today — for example, when it comes to assigning metadata. But what exactly is Artificial Intelligence, and which methods are already being used in technical writing departments? Fabienne Rothenberg and Eva-Maria Meier answered these questions in an interview with Susanne Meier from Quanos Solutions at the tekom annual conference.
Interview: Artificial Intelligence in Technical Documentation, Part 1
Ms. Lange, Ms. Wolf, the term "Artificial Intelligence" is on everyone's lips. Can you explain what Artificial Intelligence actually is and whether it's more than just a passing trend?
Eva-Maria Wolf: Artificial Intelligence (AI) is a branch of computer science that attempts to imitate human thinking and behavior. Clearly defining the term is difficult, as the boundaries between AI and complex algorithms are fluid. We take a pragmatic view of Artificial Intelligence. This means we assume that something qualifies as Artificial Intelligence "when the result delivered by a piece of software appears intelligent."
The research field has existed for a very long time. Some methods were already developed in the early 1980s. In recent years, the topic has gained momentum because greater computing power and cloud architectures have made more performance available at lower cost. This allows complex models to be processed and advances to be achieved.
In technical communication, Artificial Intelligence has been a genuine trend topic for about 2 years now. This is probably also due to the fact that more and more software products are incorporating AI.
Would you say that people are generally open to Artificial Intelligence, or do they tend to approach it with a certain skepticism?
Fabienne Rothenberg: Most people have already come into contact with AI applications in their everyday lives — consciously or not. Examples include recommendations on streaming services, voice assistants like Alexa and Siri, moderation mechanisms on social media platforms, email spam filters, and machine translation.
In the field of technical communication, the topic is not yet quite so commonplace. In our day-to-day work, we primarily encounter great interest and curiosity. However, we have also been confronted with concerns that technical writers could be replaced by AI applications. We do not see this risk materializing, for several reasons. Nevertheless — as with any new technology — there are people who are skeptical of Artificial Intelligence.
We can, however, put the fear of an AI taking over the world to rest: today's forms of Artificial Intelligence can only do what they were created and trained to do. They are always meant to serve as support — a tool that helps people and reduces tedious work.
There are different types of Artificial Intelligence. What are they, and what do they do?
Eva-Maria Wolf: The types you distinguish depend on your perspective. From our pragmatic understanding of AI, two main fields can essentially be identified: rule-based methods and machine learning.
With rule-based methods, rules are programmed to solve specific problems, following the pattern: if case "x" occurs, do "y". The rules can be combined, allowing complex systems to emerge that deliver an intelligently-appearing output. Rule-based methods also include pattern recognition in texts via regular expressions and the evaluation of knowledge graphs.
Unlike rule-based methods, machine learning does not prescribe the path to a solution. Instead, the AI learns from statistical values and experience. During programming, only the problem statement, boundary conditions, and information about how learning occurs are provided. All of this information is stored in a model. Perhaps we will return to this topic later.
The media typically refers to Deep Learning when talking about Artificial Intelligence. This is a subfield of machine learning in which complex models with many layers — consisting of numerous interconnected decision points (neurons) — are built. In principle, the goal is to replicate the human brain in order to solve complex tasks such as image recognition.
To build these many layers, a very large amount of training data is required. Companies like Google, for example, have a significant advantage here for general tasks, as they can use all images from their online cloud solutions for training.
Which criteria the Artificial Intelligence uses to make decisions is not visible. The networks function like a black box.
A cloud in which the two AI types — rule-based methods and machine learning — are each shown in their own circle. Inside the machine learning circle, there is an additional circle for deep learning, as deep learning is a subset of machine learning.
A further way to categorize AI is by the "format" of the data to be processed. In this case, one distinguishes, for example, between image recognition and classification on the one hand, and Natural Language Processing on the other. Natural Language Processing is concerned with understanding and generating natural language.
Metadata assignment is an important topic in technical documentation, and Artificial Intelligence has brought significant relief for technical writers in this area. Can you explain how metadata assignment works using so-called rule-based methods?
Fabienne Rothenberg: Simply put, rule-based methods use predefined rules to decide which metadata to assign. In most cases, a text is matched against a predefined list of values. If, for example, a specific word is found in a text and that word appears in the value list, it could be assigned as metadata. In practice, however, rule-based assignment is considerably more complex. Various rules are combined with search patterns to enable precise metadata predictions.
What matters includes, among other things, where in the text the word was found, whether the word itself or an indicator (e.g., a synonym) was found, and whether it is an exact match or a fuzzy match. In most cases, there are multiple matches in a text, and the AI uses additional rules to determine which match is most probable — or whether multiple matches should be suggested.
In addition, knowledge networks can be used to define dependencies between metadata values, or even entire product models can be set up as metadata networks. These knowledge networks can also be evaluated in a rule-based manner: if a certain value "A" has been assigned as metadata, then also assign value "B".
A simple example of this is the metadata fields "Product" and "Manufacturer". A small product model can express who manufactured the product. If the product (e.g., "T3-B") is found in the text, the manufacturer (e.g., "PI-Fan AG") can also be assigned as metadata.
Illustration of a knowledge network from which certain metadata is derived based on the network relationships.
The great advantage of rule-based methods is that no extensive training is required. Using a specific standard rule set, most metadata can be predicted. If something doesn't fit, the rules can easily be adjusted. This works best for all metadata that can be extracted from the text in some form — such as product names, serial numbers, assemblies, and activities.
Machine learning can also support metadata assignment. How does that work exactly?
Eva-Maria Meier: With machine learning, we always need — as mentioned at the outset — a model as a foundation. The model must be trained specifically for the task to be solved, e.g., to recognize the document type of PDF files.
At Quanos we work with "supervised learning" for content classification. With "supervised learning", we feed the Artificial Intelligence with labeled example documents. We need a critical mass of examples for each value from which the AI can learn to distinguish the values.
The AI extracts features from the data. These are weighted and incorporated into the model. Which features are extracted is determined by an algorithm. The weighting, however, is an adjustable parameter. The model stores the characteristic points for each metadata value in the form of a vector.
When we then show the AI unclassified texts, the same feature extraction runs and the AI compares which document type the found features best match. If the vector of the new document is close to the typical "user manual vector", for example, that value is predicted. The AI is only ever as good as the model from training. It is therefore especially important to train the AI with very well-classified examples.
That said, the AI can still be wrong sometimes. In the field, we achieve an accuracy of approx. 92% with very good models. Results therefore always need to be reviewed by humans. If the AI is wrong, a re-training with the correction can be initiated. Artificial Intelligence and technical writers work best hand in hand!
Ms. Wolf, Ms. Lange, thank you very much for this first part of the conversation!
Part 2 of the Interview on our Blog!
Use cases of Artificial Intelligence, their opportunities but also their limitations — those are the topics in the second part of this blog series. Read it here in the blog.
Do you want to learn more about the Artificial Intelligence from plusmeta or do you have a question? Then contact us here.
