Chatting with the Rest of the World: Designing Multilingual Chatbots
Even though experienced Voice User Interface (VUI) designers are hyper-sensitive to the importance of good localization over blind translation—and to the challenges of conversational automation in general—our UI team was surprised at the number of complex roadblocks we ran into when localizing a chatbot for our large international client. Read on to find out what you need to consider before undertaking a similar project.
To provide a good customer experience to the end user or “Chatter,” a global chatbot needs to be designed carefully by people with the right expertise to localize, not translate—and the skills I’m talking about are not your typical VUI strengths. We found that it was crucial to have not only experienced designers and linguists to create the conversations, but also to work with subject matter experts familiar with current local customs, demographics, and chatting norms. Who knew that we would be consulting teenagers, learning about keyboards, and counting letters?
Although people in many areas of the world now find it convenient, easy, and accessible to chat or text – even more convenient (and often cheaper) than talking on a phone – the logistics for chatting vary greatly depending on cellular and internet access in general. In some countries, texting is cheaper and easier; in others, chatting through the internet on a home computer is the norm; and in still others, talking might be cheap while WiFi is at a premium.
Any global company that wants to have an automated interaction with its customers will want to communicate with them in their preferred mode, on their preferred device. And this can fluctuate from culture to culture. For example, preliminary data from a multilingual survey showed customers in China had higher participation in text or chat than in email, and that it was very common to interact specifically on a mobile phone. For other countries, the survey responses were higher for an online link sent to the caller’s mobile phone by text. This means that a multilingual chatbot needs to be not only conversational and customer-centric, but also efficient, mindful of screen space, and aware of visual impact and readability on a cell phone. This is not the VUI we grew up with.
Screenspace and Keyboard Variations
Screen space and readability are a visual issue for customers chatting on a mobile device. In testing our chatbot in multiple languages, we found that it was universally easier to read with line breaks inserted. We also found that depending on the type of chat and locale, carriers have character limits that can kick in along with different regulations on whether or not your chatbot needs to offer commands like HELP or STOP. Some countries even have filters for specific terms or for profanity, and this can impact the conversational nature of the chats and their appearance on the screen. These types of constraints require research and local knowledge. Keeping usability in mind, we had to count the letters and condense our side of the conversation as much as possible to fit it in and make it look nice–a new dimension of VUI design for most of us. This is particularly challenging for languages that punctuate differently for commands (i.e. where capitalization is not an option and quotes or other symbols are required to make the command clear to the Chatter).
A good chatbot also needs to be aware of the physical limitations of the user’s hands and fingers. This is, of course, universal. Brevity and efficiency are at a premium in terms of how much effort the user needs to put in to respond. In some locales where callers are using devices that require a specific keyboard for their writing system, it’s a much better experience for them to respond without switching keyboards if they don’t want to. In China, pinyin is a standard system for transliterating Chinese using the Roman alphabet. It’s used by some Chatters in China for common responses like “Y” for “yes” and “N” for “no” – even though the Chatters may not otherwise use or understand the full English words. Chatbots in China commonly suggest that callers use “TD”, short for ?? (pinyin “tuiding”) meaning “unsubscribe” to make it easier on a user who wants to stop receiving chats.
After testing our chatbot, we determined that the best experience of all was to offer Chatters the option of pinyin or character responses. This way the user could choose not to change keyboards if they didn’t want to, and could answer in whatever mode was easiest for them:
The Art of Abbreviation
Meanwhile a centerpiece of chatting–which is really a written conversation–is abbreviation, which takes up less real estate on a device and saves the Chatter’s hands as much as possible. Abbreviation is accompanied by other shortcuts like lack of punctuation, capitalization, diacritics, etc.
The multilingual challenge is that the way in which people abbreviate is not the same in every locale and in every language. My daughter, who is a teenage native speaker of American English, sent me this preface to a text example from a conversation with her friend: “hey mom ily i hope u can read this bc it’s diff to understand but this is the way we txt so..”.
In this example, “ily” is a common teenage shortcut for “I love you”, and “u”/”bc” abbreviate to “you” and “because” respectively. She knows how to use commas, but doesn’t bother with them in texts and chats; and although she can write standard, prescriptively correct English for school and other communications, she is biliterate and communicates with her friends and close family in a different written chat language that not every English speaker will understand.
Chatting as a Language
Although I can translate most of my teenager’s texts, my father (an occasional texter) could not decipher pieces of her longer chats. He doesn’t know that “LOL” is now a standard abbreviation for “laugh out loud”, and he definitely would not understand the new spoken vocabulary item “lol” (pronounced “loll”) that’s derived from the chat language and that I just learned myself.
In the same way that a good traditional VUI designer aims for spoken conversations that are comfortable and intuitive for a majority of speakers, our chatbots need to do the same in the new language of chat. What’s tricky is that not all written languages use the same rules to move from standard writing to chat, and of course the new vocabulary items and other conventions will differ region by region. Spanish Chatters abbreviate differently, for example, and Spanish abbreviations in different regions of the world don’t always line up with each other either.
A screenshot of a written chat conversation in Spanish between my husband and a friend had multiple instances of “q” for the word “que” (“that” or “which”), “pa” for “para” (“for”), and also left off many (but not all) written accents. I know of no automated translation tool or algorithm that would be able to accurately localize the English or Spanish chats into another chat language, or would be able to go from full standard written form to chat (or vice versa). Even a human translator who didn’t chat the language (like my father or an ex-pat who was out of touch with the current norms) would be challenged. This adds a different layer to the process of localization; your localizer needs to speak the language, write the language, and also chat it fluently.
The Emoji Factor
While researching the chatting norms in several locales for our customer, we found that the use and enjoyment of emoticons was universal. They allow people to express an emotion or response with a symbol. But it turns out that there are different ways of creating emoticons from keyboard strokes in different regions of the world. Western emoticons (“emoji”) are different from Eastern emoticons (“kaomoji”), which originated in Japan, are often typed out using the icons, and are widely used in areas of Asia:
Western Icon (Emoji)
Eastern Icon (Kaomoji)
;-), *-), ;D
:-O, :o, >:O
( ? ??)(°?°)
So if a chatbot wants to chat with emotion, it will need to be prepared to understand the variation of emotions chatted in response.
Considering the Writing System and Localization
Yet another hurdle in localizing a global chatbot involves the type of writing system end users are used to. People in many areas of the world use writing systems that are not based on the Roman alphabet. Due to its history, American Sign Language is linguistically closer to (and more likely to be mutually intelligible by) speakers of French Sign Language than speakers of British Sign Language.
Writing systems don’t always align with their corresponding spoken languages on a one-to-one basis for the same reason. In Mainland China and Singapore, Simplified Chinese is the standard written form for both Mandarin and Cantonese speakers, even though they are not mutually intelligible when speaking. On the other hand, the Traditional Chinese writing system is commonly used in Hong Kong and Taiwan. The two writing systems can be mutually intelligible, but it’s not easy and of course, it wouldn’t be an optimal customer experience to chat with your customers using the wrong one.
For our global chatbot, we worked with an expert localizer who not only spoke both Mandarin and Cantonese, she was also a native Chatter and texter who had frequent written conversations with other Chatters in different regions. This expertise was critical to the success of our project in China.
Understand the Subtleties
There are other subtle regional differences a good global chatbot needs to be aware of, including written date formats, currency norms, sensitive political or regional terms or customs, and more. It’s not that good VUI designers can’t take all of this into consideration or wouldn’t be able to design with other cultures and dialects in mind; but if the new frontier is to chat with our end customers and if the aim is to make this interaction a good experience for them, we can’t under-estimate the importance of smart, native-chatting experts familiar with the current locale at our side.
On the last day of SpeechTEK 2018, I had the pleasure of moderating a Sunrise Session entitled "The Future of Speech-Enabled Applications." The goal of this roundtable was to brainstorm about creative and useful applications that could take conversational interactions to the next level. More than a dozen attendees from a variety of backgrounds held a lively discussion and came up with four innovative ideas, which I have summarized.