Focus
How multilingual chatbots influence localization
Kaspars Kaulins is a business development director at Tilde, a European language technology innovator and service provider offering translation, localization and custom machine translation.
Kaspars Kaulins
Kaspars Kaulins
Kaspars Kaulins
Kaspars Kaulins
Kaspars Kaulins is a business development director at Tilde, a European language technology innovator and service provider offering translation, localization and custom machine translation.
A

s technology advances, users are becoming more demanding and it is becoming harder to meet instant service expectations from customers. Companies also face the challenge of multilingualism as they are pressured to bridge language barriers and provide content in the customer’s native language. Thankfully, AI and machine learning (ML) have created opportunities for companies to deliver a consistent user experience across languages by using AI-enabled, intelligent virtual assistants. Deep machine learning and language data are two major components that can make or break these human-like interactions. Chatbots not only enable companies to serve their customers 24/7, but also create an exciting and emerging field for the localization industry.

Evolution of chatbots
We live in an era where chatbots are springing up like mushrooms after a rain. While the first chatbot was released in the 1960s, it is only in the last couple of years that this chatbot generation has reached its current state, where it can converse in a manner that feels nearly human. The advances in AI and ML, and the revolution in natural language processing (NLP) and natural language understanding (NLU), have created a generation of chatbots so advanced that often people don’t even realize they are communicating with a machine. The chatbot revolution has had an impact on all of us — from the way we live and work, to the way we interact with one another. AI has powered virtual assistants to handle complex human interaction with ease. They can help serve customers on websites, online stores and helpdesks, and automate the handling of basic user requests.

According to Gartner statistics on chatbots, experts predict that by 2021, 85% of customer interactions with companies will be handled through chatbots. Virtual assistants provide clients with instant answers and allow companies to reduce costs. Best of all, human agents don’t have to answer the same question 100 times a day, so they can spend their time solving more complex issues. The online publication Chatbots Magazine predicts that the chatbot market will continue to grow — $5 billion will be invested in chatbots by 2021. This will allow companies to cut operational costs by up to 30%. According to Juniper Research, by 2023, the use of chatbots will save both businesses and consumers 2.5 billion hours. When time is a scarce commodity, saving so much time for both businesses and customers is outstanding.

Many companies have already integrated chatbot technology, and numerous others plan to do so in the near future. No matter whether this technology is known or new, companies all face the same challenge: creating human-like assistants that provide a universal customer experience across languages. This adds an emerging field to our industry: chatbot localization.

Incorporating linguists into development
One of the most important elements that makes up a chatbot is language and its ability to converse, yet not all chatbots are created equal. A virtual assistant’s knowledge-based texts, their quality or lack of it, and data processing can make or break the customer experience. Chatbot developers not only need to invest resources in technology, but also in chatbot language development. The virtual assistant knowledge database consists of two types of texts: questions and answers.

Knowledge database development starts with crafting potential customer questions. A well-established question database gives the virtual assistant direct insight into the client’s needs and issues. Most customer service teams deal with up to 80% of the same frequently asked questions each day. These are repetitive, low-level questions that could be easily answered by an AI powered chatbot. Over time, these customer service teams have accumulated decent baseline data regarding customer needs and wishes. To train a virtual assistant to understand client needs and provide human-like interaction, it needs enough diverse training material that reflects the way customers talk and includes the many ways in which clients might express the same thought.

For example, a customer seeking help from an online banking chatbot may ask, “Where can I pay?” while another client might ask the same question in the form of a statement, such as “I want to pay” or “I need to make a payment.” To be able to process a client’s needs properly, virtual assistants need to be trained in every possible way of asking the same question. The virtual assistant’s knowledge expands further through use, as its database is constantly reviewed and improved, and through the addition of new data obtained through interactions with customers.

To compile a good knowledge database, linguists need to use their creativity and have sufficient knowledge of the customer persona in order to really fit into the skin of a regular customer. Sometimes, people don’t have a clear structure in the way they talk, and they often make spelling mistakes or grammatical errors when they are writing. The inconsistent use of different word forms, unjustified changes in expression, grammar and spelling errors can be confusing for virtual assistants and can disrupt their ability to converse if they are not trained for such situations. Therefore, creators of virtual assistants are encouraged to hire localization specialists, text editors and proofreaders to properly prepare virtual assistants for human-like interactions. For the sake of better understanding, texts that include typos and grammatical errors need to be included in the virtual assistant’s knowledge database.

The other facet of intellectual virtual assistants involves the responses that the client receives from the chatbot. To create human-like answers and converse with clients, it is crucial to develop responses in a situationally-appropriate language style, which requires additional effort from language editors as well as proofreaders and translators. Experience has shown that copying plain text from business or government websites, which frequently does have answers to a user’s question, is not a good solution for building a chatbot knowledge base. The purpose of having a virtual assistant is to provide customers with a pleasant experience. Having a virtual assistant is not enough; it must also be intelligent, with high-grade quality and the ability to adapt to different scenarios and situations.

Chatbot personality and tone
A virtual assistant’s goal is to provide the customer with short, quick and easy-to-read responses that offer an immediate and fitting solution. At the same time, it is important to consider and establish a clear chatbot identity. The language of the virtual assistant should be in line with brand values and tone of voice of the company and must clearly express its defined personality.

Depending on the target audience, companies must choose one consistent conversation method that could be either familiar or formal, scientific, professional or conversational, or they could define their own unique dialogue style. If the virtual assistant is expected to be kind, it should use courtesy phrases like “please” and “thank you.” If it is intended to break down the bureaucratic barrier between the institution and the client, the characteristics of a scientific and business language style must be dropped. One of the challenges that comes with playful chatbot personalities used in sales is the use of emojis and gifs. Not only do they create additional technical challenges regarding compliance with computer-aided translation tools, but they also require translators to be fluent in emojis and pop culture references so that conversations do not lose a personal touch or context.

Mixing machine and human translation
Machine translation (MT) can benefit companies that want to make their virtual assistant multilingual as MT can significantly speed up processes. However, it is best practice to involve a linguist who will post-edit translations for cultural references, language nuances like jokes and aphorisms, and make sure that the original text expresses the same goal as the source language. Incorrect translations can cause companies to lose potential clients, and nobody wants to deal with situations which could be easily prevented.

Often a linguist’s daily workflow is quite repetitive, and they may fall into unwanted routines. All that changed recently, when new and exciting opportunities presented themselves through chatbot localization projects. One of our linguists with more than 15 years of experience in the industry shared her insights of working on a project for Nuance, a pioneer and leader in conversational AI innovations. She also reminisced about the old days when editors used to proofread translations by printing out documents and using a red pen. At that time, she could never have imagined working with texts for virtual assistants. However, this is the beauty of technology — innovative ideas can be turned into reality, which requires an open mind and flexibility. Translators and editors need to be able to adapt and quickly and to take on new tasks and challenges. Nowadays, more and more translation projects involve dealing with AI, whether it is post-editing MT translated texts, localizing chatbots or proofreading texts generated by voice-to-text technologies.

A chatbot localization project is challenging, but also very rewarding. At first, it is quite difficult to adapt. You go from translating and editing medical or legal texts to thinking about all the possible ways how one question can be asked and answered. You need to use your creative side when crafting texts for virtual assistants, and you must understand the industry background. There are two components to chatbot localization. Firstly, creating scenarios in the source language and secondly making the chatbot multilingual.

For the Nuance project, chatbot scenarios were first created in English and then localized into the three Baltic languages. For example, there are some questions that are formally included in Latvian scenarios. However, because of the market-specific product offerings, they are not used since the products and services available in Latvia might be different to those on offer in Estonia or Lithuania.

That is the added value of involving a localization specialist as opposed to plain translation. Localization specialists consider language nuances, cultural differences, traditions and regional specificities. You can’t just translate a text directly without thinking about all the small nuances — formal word-for-word translations just don’t work when localizing chatbots. Localization specialists must use their experience to make sure that the chatbot greets customers and answers their questions in a way that is familiar and recognizable in domestic markets.

The beginning of this project was very hands-on, with endless hours per week spent with the client to create appropriate texts. As the chatbot project moved into production, the workload diminished. However, work to keep improving the virtual assistant is constant. Linguists need to review the content to see how people have interacted with the chatbots and adapt it accordingly by editing texts, adding new options and moving incorrectly answered questions to the right answer groups.

An exciting part of such projects is the ability to see the result of your work by literally interacting with a virtual assistant whose language you have helped to shape. Many times, translators return a completed project back to the client and they never see where the translation is used, but now they are literally able to interact with texts they have created. Overall, the experience of working on chatbot localization is challenging, exciting and rewarding.

Insights from an egov chatbot in Europe
Virtual assistants are mostly seen in customer service, online stores, websites and global companies; however, the public sector is also adapting these AI-driven technologies. UNA, one of the first intelligent virtual assistants in the public sector, was launched in 2018 and it was the first public administration virtual assistant in Latvia. UNA was the Register of Enterprises’ virtual employee, and this proved that virtual assistants can not only benefit enterprises, but can also help government institutions. Since then, we have developed over ten virtual assistants for various public administration institutions. We are also working on creating a unified platform for all the virtual assistants serving the public administration sector in Latvia that will be launched this year.

UNA helps clients register enterprises, liquidate companies and check documents, so human customer support can focus on the creative problem-solving that it does best. UNA has been created and trained to understand questions and respond with natural language in both Latvian and English. It incorporates multilingual NLP and NLU services for knowledge training. UNA can solve about 30% of client inquiries by itself. More complex inquiries that can’t be handled by a virtual assistant alone are passed onto the appropriate customer service representative.

A common concern is that AI might take away people’s jobs, which is why we involved the people from the Register of Enterprises to help in the development of UNA. Not only did they provide invaluable insight to make UNA better, but they also found that UNA will make their job more pleasant, instead of making it redundant. Over 64% of business respondents surveyed by Statistica believe that chatbots allow them to provide a more personalized service experience for customers. Working hand in hand with virtual assistants gives them a feeling of having a bigger impact within the company. Employees can experience new personal growth since they are now handling more complex customer issues that improve their skills. That leads us to conclude that a well-designed customer support chatbot can help a company not only to provide a better experience for clients, but can also help employees feel more satisfied in their jobs.

The future is unknown
More and more companies recognize the need for intellectual virtual assistants that can provide immediate customer support. Moreover, they see the increased value of conversational AI that can interact with the customer in his or her native language. We can all agree that chatbots are here to stay, and that they are getting increasingly intellectual and savvy. Virtual assistants will become multilingual not only in text, but also in voice, which will create a whole new set of challenges for the audio localization workflow that will affect both companies and localization specialists.

To provide the most human-like conversational AI experience, we need to include human intellect (talented translators, linguists, subject-matter experts and copyrighters) as a core value for the development of both text-based and voice-based virtual assistants. Localization of chatbot dialogue scenarios and knowledgebase requires a specific approach and skill sets, and therefore translation companies will need to develop respective competencies and capabilities. Given the growing demand and business opportunities in various industries and sectors, this is an effort well worth making.