Designing a conversational interface for payment

GoSEPP integrates its purchasing process into a German writing chatbot, and German language libraries for conversational interfaces (like a chatbot or voice assistant) are not as far advanced as those for the English language might be. Since July 3rd Anna Berger, a Cognitive Science student, analyses and designs the conversations of GoSEPP. During her Philosophy studies the philosophy and ethics of tech and AI caught her interest. Let’s explore Anna’s work a bit further.

Chatbot technology

The process of buying a product through GoSEPP is semi-automated. Conversations with users that take place after the purchase have to be handled manually. Aiming to optimize that, Anna started integrating the most common user questions using the NLP (Natural Language Processing) interface of ‚api.ai‘.

Before exploring the api.ai framework, Anna did a qualitative data analysis to find out which questions are most commonly asked after someone bought a product through GoSEPP. Mostly all users were asking questions about the same seven subjects, including solvency issues and cancelling an order.


An example of grouping all questions to one answer

Using the Api.ai library

api.ai is a framework (or: library) for building text-based conversational apps. Before starting to use api.ai, Anna tried the German beta version of wit.ai because of its ‚Stories‘ functionality, which enabled developers to create a great variety of different conversation flows. When wit.ai announced that the story function will be dropped, the switch was made to the German version of api.ai.

api.ai strongly resembles the process of human to human conversation. Every conversation needs a semantic context. To that end api.ai offers ‚intents‘, which map expressed meanings of users and actions of the software. Anna also uses the ‚entities‘ functionality, in order to extract parametric values from user input. With the entity function important data is recognized and matched. If required entities are not stated by the user, the bot uses the prompt function in order to ask for the necessary information.

Challenges integrating a NLP interface

Whilst the api.ai documentation is partly quite detailed, more examples and video tutorials would have accelerated the speed of integrating api.ai into GoSEPP.
Insufficiencies in api.ai when it comes to complex, German conversation flows, were also topic of discussion in the team. Realizing that conversational interfaces and AI are still in their infancy however, makes that GoSEPP is optimistic about being able to create more sophisticated conversations, allowing to handle eclectic dialog trees.

What’s next?

The test version of what Anna has been working on will be integrated. After a test period, data will immediately be transferred from the GoSEPP (anonymized) user chats to the training sets, to continuously optimize conversation flows.

Monday September 4, chatbot experts Thomas Schranz (Lemmings.io), Barbara Ondrisek (The Chatbot Agency) and our very own Anna will hold a panel discussion at sektor5 on the most interesting use cases for chatbots: chatbotpanel.eventbrite.com