Sign up Log in


Topic guides

How-to guides

Reference guides

Work with internationalization

8 min read.

On this page

    By analyzing sentences in different languages, it is often possible to find common patterns of how an idea is expressed. In, agents are multilingual, allowing you to leverage the full power of factoring these models together.

    Configure your agent

    The agent will be able to interpret sentences like “Pour the flour” or “Break the eggs” in English and French.

    First we need to create the agent. In the settings panel you can choose which languages it will be able to understand. Name it “Home cooking” and check “No language” and “en (English)”.

    Language settings screenshot

    Create a public interpretation named recipe_step and open it. Notice three tabs at the top, one for “No language”, another for the English language and the last one with a “+” sign. Click on it to show the “Add language” menu, then click on “fr (French)” to add this language to the agent.

    Add french language screenshot

    Multilingual agent

    Start by adding two formulations Pour the flour under the English tab and Versez la farine under the French tab. Try these two sentences in the console.

    Two formulations with naive approach screenshot

    Of course they perfectly match! The annoying part is that everything is duplicated for every language. We can do better! As you build more agents, you will see that the language dependent part can be quite limited. It is possible to factorize together the common syntactic pattern and leave language specifics words in dedicated entities lists. We are going to do exactly that.

    Create two entities lists, one called food and the second called instructions. In “food” add this entity:


    Food entities list screenshot

    In “instructions” add two entities:




    Instructions entities list screenshot

    Now, annotate both formulations with entities lists. Notice that in English and in French those two annotations uses the same pattern, namely an instruction followed by a food. It does not depend on any natural language anymore. In order to avoid this duplication we can recreate such formulation in the “No language” tab and delete the English and French formulations. This tab is the place to put formulations describing the common patterns between languages.

    Formulation factorised pattern screenshot

    If you want to add another language, simply add expressions in the corresponding entities lists.

    In this example, we are lucky enough to remove all languages dependencies in a single level, but sometimes it takes more interpretations indirections to reduce the dependency. However, one or two levels are usually enough.

    Whether you want to create a monolingual or multilingual agent, you should apply the following principles:

    • create entities with the right language: when a word come from a given language, say so!
    • add the interpretation’s words in the corresponding language tab,
    • when your interpretation does not depend on any language, add it in the “No language” tab.

    Words variations

    Words in a language are found with many variations such as plurals, verbs, adjectives, etc. In order to extend the language support of our formulations, we can use the automatic lemmatization present in

    For instance, if we try the sentence “Pouring the flour” no interpretation is found because our current agent is not able to handle such variation.

    Non matching variation screenshot

    To apply lemmatization to a word, use the following rules:

    • specify the language,
    • use singular for nouns and adjectives,
    • use infinitive for verbs.

    This means that in our entities list “instructions” we must change several things. First we must specify the corresponding language of each terms by adding the language ISO-639 norm encoding separated by a colon “:” character. Second we must change our terms to be in the infinitive forms. It is already the case for the English version but not for the French one.




    Lemmatized entities list screenshot

    If we try again the previous sentence “Pouring the flour” now it match even if we do not have it explicitly in the entities list!

    Matching variation screenshot

    NLP request and language

    Most of the time we do not need to specify what is the expected language. It is left to the NLP to match the submitted sentence against the most likely interpretation.

    Scoring with auto screenshot

    It is also possible to specify one or several languages using the “Accept-Language” parameter. In that case, all languages are accepted, but the specified language is privileged.

    Scoring with english screenshot

    This means that even if you specify “en” as the only accepted language, it will still understand all other languages.

    Scoring with another language screenshot

    The universal language

    In our everyday lives, we often use more than words to communicate, we use emojis :grinning:!

    In the “food” entities list, create a new entity for eggs in English, French and emoji (no language):


    Egg entity screenshot

    Now if you type the sentence “Break the 🥚” we have a match!

    Matching with emoji screenshot

    On this page

      Home Pricing Open Source Blog Documentation