Crossing Language and Cultural Barriers: How Does AI Achieve Multilingual Intelligence?

In the globalized business environment, few companies are confined to the geographic boundaries of a single country. Business technology platforms and tools must not only cross oceans, but also overcome language and cultural barriers.

Artificial intelligence is increasingly woven into the fabric of communication and globalization, making it crucial to equip its algorithms with multilingual and multicultural capabilities. Fortunately, teaching a computer a new language takes far less time than teaching a human, but the process is still constrained by resource limitations and the nuances within and between languages.

Intelligent multilingual systems, such as Amazon's Alexa and Google Assistant, are prominent examples of broad language capabilities in modern AI applications. As more companies replicate their success, the natural language processing community and market grow steadily.

Algorithms cannot replace Duolingo's interactive experience

Even when customers know there is no real person on the other end, they do not want to feel like they are talking to a machine when interacting with chatbots or AI-driven communication devices. This relationship must be as human and fluid as possible.

The most basic model for a system to process foreign languages is pure machine translation. As any language learner who has copied and pasted a text into a translation program knows, this process is often riddled with errors. By translating word by word and then stitching the results into a string, without translating the sentence as a whole, the result often appears fragmented.

Effective pure machine translation models are difficult to build. Researchers must establish knowledge bases that include local culture and language, Eyal Grayevsky, CEO and co-founder of Mya Systems, told CIO Dive.

Mya uses an ensemble model that compares multiple natural language understanding models against each other to obtain the best output. One of these models uses machine translation and compares its confidence with native implementations to strengthen the model over time, eventually reaching fully native levels.

Customizing solutions for any language a company wants to add is ideal, but obtaining so much data and executing it is difficult, James Maddox, CTO and co-founder of Mya, told CIO Dive.

No two languages are exactly alike

Due to unique structures and patterns, adding a new language can be difficult. Similarities within language families, such as the Romance languages, can facilitate the process of adding new language capabilities.

But when adding a language that does not belong to an existing language family, a different process may be required, Kevin Gao, CEO and co-founder of Comm100 Network Corporation, told CIO Dive.

For example, tokenization is easier in languages with space-separated words, while Japanese and Chinese have different structures.

Even in languages that use spaces, tokenization has different requirements. German is known for lengthy compound nouns, such as "Rechtsschutzversicherungsgesellschaften" (insurance companies providing legal protection) or "Geschwindigkeitsbegrenzung" (speed limit), which makes translating individual nouns difficult.

Once sentences are tokenized, words are mapped to vectors, where a series of numbers encodes their meaning. For example, the numbers for the word "brighten" are similar to "illuminate" but different from "darken," Maddox said.

Building and rolling out French, German, and Spanish capabilities each took about eight weeks, Maddox said. When Mya plans to add Asian language skills to its tools, starting with Simplified Chinese, it will require significant customization.

Google launched multilingual support for Google Assistant in August, a multi-year effort carried out in three phases: multilingual recognition, understanding, and identification.

The assistant must not only fluently understand languages including Spanish, English, Italian, German, and Japanese, but also recognize and process language pairs simultaneously, so that commands like "turn off the lights" can be understood and executed regardless of the language in which they are spoken.

Language mixing phenomena, such as Spanglish, make multilingual capabilities even more complex.

Stepping out of the data comfort zone

Obtaining data to train new language systems, just like obtaining data for any AI project, is an ongoing challenge for companies.

High-quality conversational data and natural language corpora are needed. For niche products like HR AI tools, data common in the HR domain (such as conversations about work benefits) is hard to come by off the shelf, Maddox said. Companies often improvise with "bootstrapping techniques," creating scenarios internally.

Small companies find it difficult to accomplish this alone: the challenge requires significant manpower, talent, and capital to train algorithms, Gao said. Partnering with AI partners or using publicly available online tools can help alleviate data shortages.

But feeding word translations into algorithms without cultural correspondence also results in incomplete systems.

Even within a single language, there can be vast differences. For example, in the UK, words like "fortnight" are more common, while they sound outdated in most parts of the US. And "chip" and "braces" convey different meanings in the two countries—no one wants to get chips when expecting fries, or braces when expecting suspenders. Within the US, regional dialects make uniformity even harder.

Language is largely localized and deeply tied to cultural understanding. Tech giants operating in North America or Europe must be aware of the risks of bringing cultural biases and assumptions into new markets.

This spring, Google began crowdsourcing image recognition for its machine learning systems, targeting areas under-documented beyond Western culture and affluence. For example, images unique to Asian, Latin American, or African cultures are under-documented by researchers in Western upbringing and social environments.

Algorithmic literacy beyond Western regions is crucial to the international strategy of any company seeking success in new global markets.

Entering a new language market means years of preparation. Amazon's Echo devices began rolling out in France this summer, adding to the assistant's English, German, and Japanese fluency. Amazon began using Alexa in its fulfillment centers earlier for migration, then moved to early access customers to understand common questions and identify accents, according to WIRED.

The company also developed a multilingual transliteration system for converting different language scripts, creating a database pairing names in English with Japanese, Hebrew, Arabic, and Russian.