The implementation of multilingual chatbots is not that easy. Here are a few aspects that you need to consider for implementing a multilingual chatbot.

Language detection

One of the most critical success factors of a multilingual chatbot is its ability to detect the right language. While Language Detectors i.e the translator services in the market are efficient in detecting languages and have unique scripts of their own, they have their own limitations.

It sometimes becomes very challenging for translator services to differentiate between languages with the same script. For example, the second and third most spoken languages by native speakers – Spanish and English respectively are written in the same script (Latin).

Moreover, there are around 14 languages other than English and Spanish that are written in Latin Script including French, Turkish and Portuguese.

Here is another limitation: a simple word like “No” is used not only in English but also in many other languages (Ex: Spanish). Detecting the right language becomes even more complex when users use both languages in a phrase. This is because usually nouns exist only in one language and these can be used along with any other languages. In such cases, Translator services would require more words to figure out the language. However, nowadays services like Microsoft Translator can detect more than one language for a particular set of word(s).

Language detection plays a very important role and it shouldn’t be just dependent on what a translator service detects. If the language detection has to be efficient, your chatbot solution should also be able to influence the detection based on the user’s geography, the context of the conversation and other user details if available through user profile or user AD.

Language understanding

Though Chatbots, with minimum training, can figure out the user intentions for most of the languages – one of the main challenges chatbots face these days is the ability to understand the different regional accents or linguistic varieties (Dialects).

Training the language understanding model with patterns of different linguistic varieties can help mitigate this. But if you are looking to roll out your chatbots to multiple countries in multiple languages, you might want to consider Language Understanding services that support different local languages.

A Language understanding model equipped with quality training data is second to none. But if your Language Understanding service provider can offer support in multiple locales, it’s even better.

Because the efficiency of language understanding will be better if the service itself supports the locale.

For instance, LUIS, the language understanding service of Microsoft supports both France’s French and Canadian French. So, if you are rolling out a chatbot in France and Canada with language support for both English and French, you might want to consider having different language models for English, France’s French, and Canadian French. This will help the chatbot in understanding better and will reduce the need for larger training data.

Another important aspect of Language Understanding when it comes to multilingual chatbots is Entity training. Although users tend to write nouns mostly in their respective languages, it is not the case every time. For instance, while some words are spoken in English, they have their own language variations as well.

For example, the Russian word for ‘Manager’ is Менеджер. So, when you are training entities in your language model, don’t assume that all nouns to be in the same language –  consider the language variations as well.

Conversations

It is not just the capabilities of a multilingual chatbot that make it successful – user experience counts as well. Every chatbots’ user experience relied on the design of its conversations. One of the typical strategies used when rolling out chatbots in multiple countries and languages is building the chatbot personality at a global level. This is not only incorrect but also risky for a chatbot roll out.

Cultures differ with regions. Some conversations that are polite in one country aren’t deemed the same way in another country. The word “Crazy” might sound funny in the UK but it’s offensive in the US. So, it’s very important for Conversational Architects to build a chatbot personality at a country level than at a global level.

This means having a single Conversational Architect for a multilingual chatbot wouldn’t be enough – even if he/she has an exceptional cross-cultural understanding. Having a cross-cultural team of Conversational Designers is a better bet. This will help to bring in the flair of language in conversations which might be very locale-specific at times.

One another important aspect of a chatbot conversation is Context Management. For example, the word “Ciao” in the Italian Language is used both for greeting and farewell. A chatbot with poor context management will not be able to differentiate scenarios like these. Thus, it is important that the chatbot has the capability to understand the state of the conversation and act accordingly.