Frequently Asked Questions (FAQs) are among the first places users go when they need information. The advantage of FAQs is that they are confined to the most important points and most common problems. This means it is much easier to monitor your FAQs than your entire documentation. But to facilitate this, FAQs must be well sorted, easy to navigate and, above all, up-to-date. In other words, the success of FAQs stands and falls with the capabilities of the editorial team. A large number of questions results in a great deal of work for the FAQ editorial team. However, a high volume and frequency of individual questions can also be advantageous. The “F” (“Frequently”) in FAQ enables automation that minimises manual and / or repetitive editorial work. Here are a few ideas on where to start.
Automatic Recognition of Similar Questions
A first step towards FAQ automation is recognising similar questions. If users cannot solve their question (quickly enough) via FAQs or navigation, they will switch to other channels. Most of the time, this is either a search or a service centre. If the Service Centre is contacted via ticket, email, or chat, or if requests are logged, the collection of user questions expands. Depending on the format and source, this collection can be automatically and continuously searched for patterns.
Identifying common terms in search queries is quite simple to do. More complex methods are needed to identify common themes and similarities, e.g. in the problem descriptions in a ticket. These methods come from the field of Natural Language Processing (NLP). However, using topic modelling, distance measurement and dendrograms, it is still possible to recognise groups and patterns.
If a topic is identified as occurring particularly frequently, the system can flag this for the FAQ editorial team. The system will list the questions and searches that are associated with the topic. Instead of going through all the questions themselves, the editorial team only have to decide whether the detected pattern is meaningful and relevant.
Automatic Detection of Duplicates
Another step towards more automation of FAQs is the detection of duplicates. If the same question has two different answers, this is not helpful for users in cases of doubt. On the contrary, it leads to confusion and frustration. That’s why duplicates must be detected and eliminated. Before adding a question to the FAQs, it is important to check whether it already exists.
With good sorting and few questions, this should not take too much effort. But for complex areas with different subdivisions and perspectives, it’s a different story. And if several FAQs need to be merged or different departments handle their own FAQs, this can also cause complications.
Recognising duplicates works similarly to recognising similar new questions. In this case, the difficulty is that the texts are very short, which is a problem for many NLP techniques. In addition, there is little data with which to train an intelligent system. This is because in most cases duplicates are simply removed, rather than being marked as such. Listing questions with similar keywords and categories is therefore the more efficient approach. Then the editorial team can make comparisons with the new question with minimal effort.
Another possibility for FAQ automation is in the area of sorting. In order to keep the FAQs clear or make them more accessible for a search, the questions are almost always divided into categories or even arranged in a topic hierarchy. This sorting can also take place automatically. To do this, the AI learns the common features of the questions in a category or branch of the hierarchy. It then checks which group the new question is most likely to fit into. If a group becomes too large, the system can also examine which questions within a group are most thematically similar and thus suggest a subdivision.
Furthermore, if subject areas and organisations change over time, the old sorting criteria may no longer fit. A new structure can be created based on similarities between questions. However, an editorial follow-up is strongly recommended. This is due to the fact that, while the results have the advantage of being aligned to real data, independent of personal preferences, the subdivisions are not necessarily comprehensible to a human reader.
AI and NLP can support the automation of FAQs in many places. They reduce effort and help with uniform structuring. This gives the editorial team the opportunity to concentrate on content and design work and thus continuously increase the usefulness of the FAQs and thus user satisfaction.