Skip to main content

Why do businesses need multilingual bots and how to build them?

machine learning
(Image credit: Image Credit: Geralt / Pixabay)

At the recent voice tech event Conversations V conference, Brielle Nikoloff, co-founder and product director of Botmock, presented the key points to be considered when developing a multilingual bot. We gathered the top insights from Brielle’s speech and supplemented them with our own expertise. 

Why do you even need a multilingual bot?

Multilingual bots support requests in three or more languages. Unlike the bilingual and monolingual bots, that understand only one or two languages. The functionality of multilingual bots depends on the purpose of automation: answering user requests 24/7, providing helpful information, receiving and placing orders, helping to choose the right product, etc. But nuances may appear when adapting this functionality for different language audiences. 

Multilingual bots are perfect for large companies operating in several countries or global social projects. For example, in response to the coronavirus pandemic, UNICEF/ECARO and WHO experts launched HealthBuddy Europe, an interactive chatbot that helps to obtain accurate Covid-19 information. The bot is currently available in 19 languages, including English, Spanish, Italian, Kazakh, Portuguese and Russian.

Multilingual bots are essential when users in a company's region of presence speak different languages. For example, according to KPMG and Google, in India, 73 percent of users speak Hindi and 27 percent speak English. But there are also other languages and numerous dialects. Given this fact, the Delhi-based Dr. Lal PathLabs Diagnostic Medical Center decided to implement a multilingual bot in its customer service.

How to do it

Different NLUs (Natural Language Understanding) may be used to implement the multilingual bot. The supported languages and the modes used by the bot to switch from one language to another vary pretty much and depend on the capabilities of the platforms (Botmock, Xenioo, BotMyWork. etc.). For example, Dialogflow supports over 50 languages, and the bot picks the language depending on the device or channel location.

Thanks to CAILA's (Conversational AI Linguistic Assistant) built-in multi-language NLU core JAICP (Just AI Conversational Platform) can be used to create bots in 16 languages. Also, external NLU services may be connected to the project to support other languages. The communication language can be set by the channel settings, while the platform also allows for dynamic language selection during the conversation. That’s how a bilingual bot works: for instance, a client is communicating with a Spanish-language bot, and then writes a request in English. The bot switches the context to the English-language bot – so now the client continues communicating with that bot. Later the user can write a request in Spanish again, and the context of the conversation will be returned to the Spanish-speaking bot. Notice that the scheme described above will not work if you try to add French to Spanish and English: switching contexts too often with no returning can lead to bot stack overflow. To solve it one more component must be added: the router. The bot router is an intermediate link between monolingual bots, it detects the language of incoming requests and redirects them to the right bot.

What to pay attention to 

Semantic differences in words 

Semantics as a linguistic science studies the semantic meaning of language units, such as words and expressions. When creating multilingual bots, you cannot translate reflexively, because the meanings of words in different languages may be different.

Translation level

A translator interprets the meaning of a text in one language and creates an equivalent text in another. The main purpose is to convey the meaning of the original without distortion. There is a type of translation that requires a higher skill and creative approach - transcreation. It is used in advertising or marketing, it’s used to translate film titles. A specialist who is engaged in transcreation must be focused on the new target audience and make the text as transparent and well-understood as possible. Transcreation teaches us to think about the situation as a whole, to take into account the differences in user psychology and language features, so this approach is really useful when designing multilingual bots. 

Cultural differences

Sufficient cultural differences between users of a multilingual chatbot mean that the identity of the bot needs to be re-created from scratch for each market. For example, there is a game with a voice bot made in the form of a good-natured letter carrier. A Frenchman would be delighted if the virtual character asked how his daughter's birthday went. An Englishman would be wary. He'd think, ‘How does that strange letter carrier know it was my daughter's birthday yesterday?’

In one country, small talk is inescapable before getting down to business. In another region, they don’t want to waste their time. In China, it is considered indecent to say ‘no’ openly. Cultural characteristics influence the nature of the bot, the vocabulary, structure of questions and answers, the purpose and structure of the user experience. The bot script should reflect all of it.

Team composition

In order not to miscalculate the translation and the psychology, get a native speaker on the team. This is especially true for complex languages like Chinese and closed cultures like Japan.

Testing process 

Test with native speakers and those users who speak the bot’s language as a second/third language. Theoretically, there could be a situation: during a trip abroad, a Spanish-speaking client would want to use the bot and choose English, if there is no option with his native language.

Anna Prist, writer and tech evangelist, Just AI

Anna Prist, writer and tech evangelist at Just AI.