Campus Support

FABi is trained to answer questions from the areas of "interest in studying", "application", "enrolment" and "everyday life at university" - in other words, it is aimed at prospective students and students. It is therefore integrated into the university's central pages aimed at these target groups, such as here: "Starting your studies". These subject areas are particularly suitable for the use of a chatbot because there are many simple questions here that do not require any special context (as would be the case with exam topics, for example), because many interested parties have similar questions and because a short dialogue is often sufficient to help the person concerned.

In terms of subject matter, it is limited to simple organisational questions that are typical for initial contact: "Can I study mechanical engineering at Bielefeld University?", "When does the summer semester start?", "wWe do applications?", or even "Does psychology have an NC?"

The chatbot FABi is currently being trialled as part of a project aimed at establishing standardised and transparent support structures in the area of studying and teaching. In this project, a working group is responsible for the maintenance, strategic orientation and further development of the bot. It is important for us to have the perspective and expertise of various advisory departments. Employees from these areas are represented in the working group:

• Department SL: Teaching and Learning Services department
• Centre for Lehren und Lernen [Centre for teaching and learning]: Central student counselling service
• Communication Office: Digital and Online
• BITS: Information system and process support (advisory)

It is very good at answering simple questions (such as "where can I find ...", "can I study mechanical engineering ...", "when does the application period end ...") and conducting short dialogues about them.

It is very well suited to helping with searches in existing information pages by presenting various topics to choose from or recognising a topic from the user's question and then referring to the appropriate websites. It is preferable to a Google search (which is often used to search university websites) because we specify in the chatbot which websites we offer when certain questions are asked.

It is also well suited to refer to the appropriate advice centre on a specific topic, either directly or at the end of a dialogue.

It "learns" insofar as we can train it better and better with a growing stock of example sentences and real dialogues with people seeking advice so that it recognises what the intention of a question is. However, he cannot learn new answers himself or ask questions that he has not been taught verbatim.

The chatbot cannot replace advice from case workers or student counselling services, nor can it conduct longer dialogues that would be equivalent to a real counselling session

In the version we use, it cannot search ("crawl") websites itself and learn from them, or independently gather information from an existing knowledge base. For example, the chatbot cannot answer the question of whether you can study a certain subject ("English Studies") at Bielefeld University by accessing the course information itself; however, it does refer to the corresponding page in the course information where the courses offered are listed.

This distinguishes it from bots based on "large language models" (LLMs) such as Chat GPT, which from the user's point of view are characterised above all by the fact that they generate human-like texts and react both to their own knowledge base and to user input. This enables a form of "conversation" that is similar to communication between human dialogue partners.

Various departments at the university maintain websites on the topics of application, enrolment, study organisation and advice, and there are support email addresses, contact forms and telephone numbers. So the information you are looking for is all "there", you "just" have to find it.

When in doubt, "googling" for an answer is the quicker and easier way to find a solution than navigating through a large and complex collection of websites. From an organisational point of view, this can be problematic because we have little influence over what information and which website Google displays. The first search result is not always the best result, some sources are also outside the organisation, are outdated, or are websites that fit the question but which we would not recommend. With a manually maintained chatbot, we can enter changing information directly and control which websites are referred to ourselves. We therefore have it in our own hands which answer the person seeking advice receives.

LLMs like Chat-GPT seem to know a lot, because you can ask them anything and get a comprehensible answer that seems to make sense. The factual knowledge of LLMs is very broad because they are trained with very large data sets and thus acquire their knowledge. However, this knowledge has limits: Not all of the knowledge presented in training is retained, much is discarded - the model learns "on gaps", so to speak. And the knowledge of the LLM has a date - the model does not recognise information that was created after the degree of training. Nevertheless, it gives seemingly plausible answers - even if it only knows the facts "approximately".

FABi's answers can seem a bit "wooden" compared to an LLM and they are always the same, a conversation that seems like a natural exchange is not possible (or only for a short time). But FABi sticks to the facts in his answers. It is not possible for him to give answers other than the ones we have taught him.

If we as a university are asked when the enrolment period ends, whether you can study mechanical engineering in Bielefeld and how many credit points a bachelor's degree has, then we are not allowed to give a plausible-sounding answer that roughly fits - all advisory centres have the right and the obligation to provide correct help desks. A chatbot like FABi is part of communication management and, just like employees in support facilities or advice centres, must either give correct answers themselves, present the link to the appropriate website or refer to an advice centre. FABi can do all of this.