The role of language in the construction of a Conversational Interface

Kishor Nagula

As mentioned in the primer, using CIs to interact with customers through freeform African languages can be challenging given the limited amount of localized digital content to source.

Of the conversational interfaces we tested, it was consistently challenging to have a human-like conversation with the interface. Although the CI generally understood basic commands that follow standardized spelling and grammar conventions, it often could not understand or follow the cadence of a human conversation. Training a CI to understand syntax, sentiment, and jargon in languages that are well represented online requires sourcing enough content to understand different scenarios, which is both time-consuming and challenging. Given how new these technologies are in emerging markets, it is possible that the technical teams were not familiar with or did not fully understand how to build complicated natural language processing engines.

This becomes considerably more complex in emerging markets where either the user mixes words from their local language into their English-based conversation, or chats with the CI entirely in their local language. Moreover, the limited amount of content in local languages online makes sourcing the requisite content for the CI difficult, if not impossible, and renders the local language interaction ineffective. This means that, until there is a stronger presence of underrepresented languages online, it will likely be impossible to build in native languages in sub-Saharan Africa.

Example 1: Using buttons and dropdown menus to interact with Simply’s conversational interface

To account for these challenges, we have seen conversational interfaces structure the interaction around a combination of natural language, dropdown menus and image based buttons. As seen in the screenshot from Simply’s CI, the user can interact with the interface by using buttons or selecting options on the dropdown menu

In these cases, the natural language processing engine is used for very basic commands that are well understood and require an almost binary response, while the buttons and dropdowns are used for more complicated responses. Using these two features limits how much nuance the user can exercise in responding to a CI and ensures that the CI is able to provide an immediate and accurate response to queries that could otherwise get lost in trying to understand syntax.

Sign up for notifications of future blogs in this series