The Periodic Table of Artificial Intelligence
TL;DR Artificial Intelligence comes in many different flavors and it can be quite difficult to identify and evaluate all the ingredients in an AI solution. This is why we like the periodic table of artificial intelligence. It highlights the various ingredients and clearly explains areas of application and benefits. In this blog post we highlight the various AI elements that give semantha her capabilities. Let’s dive down into the engine room and we’ll take a closer look at one or the other cog!
Note: This blog post is about the Periodic Table of Artificial Intelligence [1] – a means for explaining the different facets of AI to everybody, especially decision makers. It is being developed by Bitkom, Germany’s digital association. Founded 1999 in Berlin, Bitkom represents more than 2,700 companies in the digital economy: more than 1,000 SMEs, over 500 startups, and virtually all global players are members. Bitkom is pretty German 🙂 – so , unfortunately, their publication “Shaping digitization with the periodic table of artificial intelligence: A navigation system for decision makers” [1] is available in German only. Read on if you’re interested in a short overview of their efforts.
Artificial Intelligence – (merely) a Collection of Things
What is Artificial Intelligence (AI)? There is currently a huge hype about what AI is (or should be) and what it can do (or should be able to). Depending on the point of view, you could say that I need AI if I want to do something that a computer cannot do (as of) today. But that’s almost philosophical – and we’re technicians. So: AI is an umbrella term for many different technologies that aim to draw conclusions from data that appear intelligent. Games (e.g. chess, Super Mario and Go), interacting (e.g. autonomous driving) and analyzing (e.g. fraud detection, customer segmentation etc.) are only applications – but different AI modules are often required to solve a single problem. Bitkom describes exactly these building blocks – regardless of technology used – in the periodic table as AI elements [1]: Yet, the periodic table does not define AI, but rather provides a framework and gives us a well-defined vocabulary [2]. Thus, we can use the AI elements to describe what semantha does and what benefits companies can derive from employing her [3].
The AI Elements of semantha
Boiling it down to the simplest, semantha processes unstructured documents in three steps:
- Step 1 determines semantic similarities between text passages.
- Step 2 extracts data points from text passages and thereby structures them.
- Step 3 links data points with background knowledge to draw logical conclusions.
Before analyzing documents, semantha must read them and transfer them into an internal data structure. If the input document is not only a plain text, but a – say – visually appealing document (e.g., a PDF document), semantha not only accesses the textual content, but also uses the visual properties of the document (image recognition, Ir). For example, tables are recognized and evaluated separately.
Then semantha compares text passages with each other and checks for overlapping content (language comprehension, Lu). The special thing about semantha’s language module is its independence from wording: it captures texts on the level of meaning. Thus, the actual wording used is less important. Semantha identifies text passages that have the same meaning or that are close together, for example to check whether a document contains predefined hotspots. You can also compare two documents (or two versions of a document) directly with one another.
Regardless of the specific application, I’d like to highlight one thing in particular: semantha comes with a given understanding of language (we call that “pretrained”). With it, she can work on tasks right from the start without the need for application-specific or customer-specific training. Her language module is by no means fixed, but can be adapted using various methods (Knowledge Refinement, Lt). In the simplest case this is an update of her configuration. But depending on the technical language / jargon, we can individualize it based on customer documents. semantha then has a tailor-made language module.
One use case for semantha is a 1:n document comparison in which the document is compared with n others. For example, you can easily compare your own terms and conditions with those of competitors: semantha then determines those text passages from competitors that match or are similar in terms of content to your own regulations. This 1:n comparison can be generalized to an n:m comparison. This allows the user to explore the semantic categories that are being created implicitly (Category Learning, Lc) and save them for future analyses.
In the second step, specific data points can be extracted – either based on the results of the similarity analysis or on the entire document (Text Extraction, Te). Both classic methods and machine learning methods are used here to obtain typed data from the text passages.
In the third step, the semantic interpretation (or reasoning) generates new information from the data obtained, such as recommendations for action together with a reason (Synthetic Reasoning, Sy). Machine learning is not used here, but the use case is modeled accordingly (here again: Knowledge Refinement, Lt). Modeling is the only way to logical reasoning. Here semantha also draws on external knowledge that is not explicitly in the document, but can be derived and/or linked using logical reasoning. semantha not only delivers the result, but also the logical chain that led to the result. In doing so, she always refers exactly to the point in the document on which her conclusion is based. This enables the user to understand her decision (we then speak of explainable AI [4] or XAI).
An Example from the Insurance Industry
In insurance underwriting, brokers draft extensive insurance conditions that they offer their customers as a special concept. Insurance companies, in turn, check whether they want to accept the risks (and at what price). Insurance companies therefore have to check the conditions proposed by the brokers. (See also our blog post “Accelerate Time-To-Market with AI”.) Experts have to read the draft clause-by-clause and examine whether the internal underwriting guidelines are being met or at which point the proposal violates them (and how).
Let’s consider the following example. The insurer is ready to accept liability claims under the German Environmental Damage Act (Umweltschadensgesetz). However, the guidelines stipulate an upper limit of € 5 million per loss event. Underwriters must therefore identify the regulation(s) on environmental damage in the terms proposed by the broker, determine the value of the sum insured for each damage event and check it against the guidelines.
Exactly this process can also be carried out (or supported) by semantha: In the first step, the (or rather: all) corresponding clause(s) on environmental damage is/are being determined. Then, in the second step, the sum insured is extracted and, in the third step, compared with the internal limit of € 5 million. If the amount insured is below the specified amount, the passage can be marked green – otherwise red. The check at the end could of course also take into account other data points, such as the information whether or not pollution of rivers (in contrast to stagnant waters) should be covered.
The last check can be carried out by semantha, since the data extraction extracts typed data. This means that the character string “€ 2 million” can be extracted as a monetary amount and then has the same value as “€ 2,000,000.00” and “EUR 2M”. Without typed extraction, we would have three different extracts. With typed data, we can, e.g., determine which amount is lower and which one is higher.
Many Possibilities – One Platform
As we have seen, semantha has a wide range of functions and uses various AI elements. But of course not every use case requires all the functions and all elements. Therefore, semantha supports various processes directly with user interfaces and guides the user through the necessary analysis steps. In some cases, we can even completely hide the analyses and show only the result and the reason to the end-user. And of course she supports the common file formats in the respective domain of application (e.g. ReqIF for requirements engineering).
We cannot foresee all of the customer’s processes. That is why we make all functions of semantha available as a REST API – of course with extensive documentation. This simplifies the integration of semantha into existing IT landscapes and process leading systems.
[1] Designing digitization with the periodic table of artificial intelligence: a navigation system for decision-makers. Original (in German): Digitalisierung gestalten mit dem Periodensystem der Künstlichen Intelligenz: Ein Navigationssystem für Entscheider; Bitkom, Bundesverband Informationswirtschaft, Telekommunikation und neue Medien e.V. (Herausgeber) https://www.bitkom.org/sites/default/files/2018-12/181204_LF_Periodensystem_online_0.pdf
[2] The periodic table of AI. Original (in German): Torsten Hartmann und Stefan Holtl: Das Periodensystem der Künstlichen Intelligenz. – Big-Data.AI Summit 2018, https://youtu.be/N3E5L2aSZkM?t=421
[3] Bitkom: Periodic Table of AI Explains Artificial Intelligence, press release. Original (in German); Bitkom: Periodensystem der KI erklärt Künstliche Intelligenz, Pressemitteilung vom 10. April 2019. https://www.bitkom.org/Presse/Presseinformation/Periodensystem-der-KI-erklaert-Kuenstliche-Intelligenz
[4] One can also discuss whether semantha is an interpretable AI or an explainable AI. As with the AI elements, it’s a mix here too. The language understanding of semantha itself is “only” explainable. Everything that builds on it – in particular the extraction and the reasoning – are interpretable, as we not only understand the underlying modeling but can even carry out the processing steps ourselves step by step with pen and paper. More details about AI vs. explainable AI vs. interpretable AI is available, for example, in the paper by Gilpin et al. Explaining Explanations: An Overview of Interpretability of Machine Learning.