AI Showcase: Simplify communication intelligently with your own AI model

This AI showcase demonstrates the possibilities of a self-contained AI system in companies and authorities. It shows how employees can be relieved with AI. For example, complaints, citizen inquiries, support tickets, emails or damage reports can be recognized and processed purposefully. Features: full data control, full independence, no additional costs.

Introduction

In the previous AI-Showcase, it was demonstrated how impressive images can be generated on an AI-laptop with little effort and at lightning speed.

In this AI showcase, it is shown that with almost equally low programming effort and within minutes an intelligent classifier for documents can be trained. This allows communication to be significantly simplified in companies and authorities.

Requests from citizens or customers can be sent directly to the right handler or automatically assigned to a suitable category. Even recommendations for the best answer, based on previous comparable cases, are possible. All this automatically and AI-based. Work processes supported by AI instead of replacing work processes. Because Reliability is based on a prerequisite: Responsibility. And this can only be taken over by a human.

An intelligent classifier can understand the meaning of documents or requests. The possibilities that follow are enormous. The following process shows in bullet points what is possible in your company or in a government agency:

• A customer or citizen writes you a message. This can be a damage report, a complaint, an appointment request, a termination, a question about something, or anything else.
• One of your employees should edit the message.
• With the intelligent classifier, it can be recognized how a customer's or citizen's message is to be processed.
• This leads to a series of possibilities:
  • The message will be automatically forwarded to the right employee.
  • The message is compared with earlier messages. An employee receives a recommendation for processing based on an earlier case that was very similar.

An alternative is a ticket system for support (customer support, citizen service, programmers) conceivable: :

• A new case comes in, for example, a problem report or complaint or a software error.
• With the intelligent classifier, this case can be automatically sorted into a suitable category.
• The right employee receives the case for processing.
• He optionally receives a recommendation on how a very similar case was handled in the past.
• The reporter (customer, citizen, user) receives a response that is appropriate for their request: "Dear customer, thank you for your message. Your message has been classified as an error report and will be processed with high priority. One of our employees has already been informed."

The Showcase

Work is done with an open-source AI model that understands German well. While the whole world speaks other languages besides German, at least AI language models understand our language very well.

The goal is to determine the category of a given document. A document is equivalent to a piece of text. This can also include an email, a message from a contact form, a web search, or input into a chatbot. This showcase works very well with texts that are no longer than a few hundred words. For longer texts, the approach used can be refined with conventional methods.

You determine which categories exist. For example, if there are employees who handle complaints on a specific topic, a category would be "Complaints on topic X". Another category could be "General Inquiry" or "Purchase Interest" or "Question about getting an ID card".

For these categories you have defined, example documents are now needed. Often, just a few examples per category suffice.

If there are very few examples, they can be multiplied by a AI process. One speaks here of synthetic datasets. For example, a new version can be generated from an existing document that is linguistically different but belongs to the same category. In many cases, conventional programming can already multiply the examples collected by humans.

Since your documents are your own, it's often important to have Data Control. That's why the AI process runs locally on your hardware. For this showcase, a AI laptop was used where this text is also being written. Companies would use their (own or rented) AI server. What's possible on a laptop is even more possible on a box called Server.

Then begins the exciting part, the training of a classifier. This classifier learns when a document is to be assigned to a certain category. How does it learn that? Exactly as if you were to look at hundreds or thousands of documents and learn to recognize the respective category yourself. The difference is that you neither have the time nor the desire nor the good memory nor are fast enough.

Results

The training took only a few minutes on the aforementioned AI laptop:

Training runtime: 212.2422 seconds

Training samples processed per second: 155.671

train_steps_per_second: 9.729

Total optimization steps = 5443628

One of the outputs of the AI training process

This output came out after a first test run. A refinement of the result took a bit longer than the mentioned 212 seconds, which is less than 4 minutes calculation time. In this time approximately 5.5 million optimization steps were performed. On a laptop (the fan is sometimes a bit loud).

Approximately ten thousand training steps were executed per second. Similarly, over 155,000 examples were hammered into the AI brain per second so that it learns to understand. The following image shows the statistics from another AI training. It's nice to see the staircases indicating the success of the training. The learning rate goes down with time (here intentionally) just like in humans (where it is often not intended).

The AI classifier can now classify documents into predefined categories, which it does quite well. In the picture, the value 0.0001 is labeled as "Loss". The "Loss" is the very small gap to perfection here. A value of 0 is perfection. A value like 0.0001 should be considered perfect. The quality of the system can only be improved further by additional training data or validation data.

For example, if the classifier is trained on user questions as they are entered into search engines or virtual assistants (simple chatbots) embedded on websites, then the AI knows what kind of user question was entered. .

Here are a few examples of user questions along with recognized categories (spelling errors are intentional):

• Hallo → Category „Social Communication“
• What is your or her name, dear bot? → Category „Personal Question“
• What day is today? → Category „Question about Date“
• Wo finden ich denn eigentlich Infos dazu wie man Personalausweise beantragen tut? → Category „Question about Identity Documents“
• programmfehler termintoll → Category „Error for Third-Level Support“
• Hilfe Who can help me fill out this formula for surrounding an automobile, please? thank you dear sir or madam → Category „Question about car re-registration“

The AI classifier can handle questions with spelling errors, incorrect terms, and sloppy or colloquial phrasing.

Further and completely different examples are not only conceivable but also realizable. Either with your own AI system, or via interfaces to commercial chatbots.

Advantages

The intelligent classifier can be immediately adapted to specific circumstances (news or documents and their respective categories) after it has been set up once.

Once available, the intelligent classifier can tirelessly complete its work and deliver fast answers.

If the desire arises over time to have new or different categories, the classifier is retrained. This only takes a few minutes again, or a few hours in the case of many examples. This effort is a one-time effort.

Try Offline-AI now

Optimizable and with full data control. Economical even in continuous operation.
Fully-controlled data center, no third-parties.

Try now

If everything is done on a own AI system, the classifier can be improved infinitely often. This is important because each training process is subject to chance. It's roughly similar to when a person writes a text on the same topic on different days (each time a different, content-wise comparable text would arise).

Furthermore, no data and valuable insights are flowing to third parties if you don't want them to. Those who are already ready to use third-party services should not only think about the API of ChatGPT. There are now offers from other providers, also from Europe, which above all deliver better results for German texts.

The economic feasibility of this solution is enormous. The benefit is still there and the effort is minimal.

Combining this AI approach with conventional methods further increases the reliability of the results.

Conclusion

Based on a few examples, an intelligent classifier for documents in companies and authorities can be created. The creation takes place in the form of targeted training of a AI model.

The hardware requirements for this training are manageable and can also be met with relatively low-performance rental hardware.

The effort primarily consists of providing a few examples. The existing examples can be artificially multiplied. The programming effort for the AI mechanism is low.

In a subsequent AI showcase, it will be demonstrated how documents can not only be classified but also compared with each other. This way, even more processes in the company can be made even more efficient. More efficient means here that the existing employees can do their job better and due to the relief from stupid tasks will probably be happier. The number of jobs will remain the same thanks to this AI relief.

We would be happy to provide an assessment of your use case in your organization regarding feasibility with AI. Just write to us.