It’s a popular story: contact centers replacing human agents with chatbots and artificial intelligence. Sinitic is one of the companies at the forefront of applying natural language processing and deep-learning theory to contact center automation for some of the toughest non-English languages, such as Thai, Chinese and Japanese.
However, for the foreseeable future, the predictions of near-total human job replacement are overblown. It is an absolute reality that chatbots will not be able to answer every possible query, and human agents will need to be on stand-by to takeover complex conversations.
It is also a reality that most multilingual contact centers cannot bear the cost of having 24/7 x 365 shifts of native-speaking agents for every language. And in most cases, if a conversation in a rare language requires a human agent during an off-hour, the customer will either be asked to come back later or submit a ticket and wait for a reply.
There are contact centers that consider this to be an acceptable outcome; however, these businesses are failing to deliver a competitive customer experience in a global consumer market that is increasingly expecting immediate and localized service. These businesses accept this outcome at risk to their customer retention and reputation.
If we take the consumer perspective, this situation isn’t hard to understand. In our personal lives, don’t we all get annoyed when companies make us — their busy, paying customers — wait? If you have ever been on an excessive hold or told to come back later, it is likely because the management of your chosen service has accepted this dismal customer experience standard in the name of cost savings.
Perhaps you should look for a different service — a service that offers immediate support at any hour in your native language. But how would these contact centers offer such an awesome service without the massive, prohibitive payroll?
Before going further, I know what you’re thinking: Aren’t machine translations terrible? Don’t they fail to understand unique terminology, typos, and slang?
In short, yes. While machine translation to-and-from English has improved in the last decade, machine translation between non-English languages is still elementary and unsuitable for the enterprise contact center environment. Here’s an example comparing Thai-to-English and Thai-to-Japanese:
Thai: คุณอยากไปเที่ยวที่สิงคโปร์มั้ย (Would you like to travel to Singapore?)
English: Would you like to travel to Singapore?
Japanese: シンガポールへ旅行しますか (Do you travel to Singapore?)
Thai: คุณกับผมรู้จักกันมานานแค่ไหนแล้ว (How long do we know each other?)
English: How long do you know me?
Japanese: いつから私を知ってるの (When did you know me?)
You can read that the Thai intent is delivered in English; however, there is an unacceptable possibility for intent confusion in Japanese. This scaterted intent is often caused by grammatical differences that are not translated properly between non-English languages, causing an over-reliance on exact word-to-word translation.
Compounding this problem is vocabulary from specific domains. Machine translation services use general datasets, so unique terminology and slang poses a barrier to the clear delivery of intent. Here’s an example:
Thai: ร้านข้าวที่ใกล้ที่สุดอยู่ที่ไหน (Where is the nearest restaurant?)
English: Where is the nearest rice shop?
Japanese: 一番近いお米屋はどこですか (Where is the nearest rice shop?)
In this example, ร้านข้าว (restaurant) is a Thai compound word which consists of two words: ร้าน (shop) and “ข้าว (rice). Since “ร้านข้าว” is used in casual conversation, the machine translation fails to understand its exact meaning, and instead translates exact words — this is the source of confusion.
So, how is machine translation supposed to be the solution?
It is more accurate to say that it is apart of the solution.
Sinitic encountered this problem with rare language translations last year, and has since been developing a Neural Machine Translation (NMT) technique that allows continuous re-training of machine translation models for non-English languages in niche domains. Literally, this technique could be called 'adaptive translations'.
Let’s look at an example. Within the Online Gaming domain, we used chat history and a deep-learning algorithm to generate a Thai-English parallel corpus (i.e. a set of translations). Native-speakers continuously corrected the most representative translations within this corpus to the extent that the Bilingual Evaluation Understudy (BLEU) score for Thai-English ranked 29.61 compared to 21.36 from common machine translation service, Microsoft Azure.
However, the BLEU scoring is an imperfect analysis. A direct comparison of translation pairs shows the capacity for adaptive translations to assimilate typos and domain-specific terminology:
ไม่ได้รับเคดิต (I don’t get the credit)
Microsoft: No K-Debit
Sinitic: Don’t get credit
ต้องการฝากเงินคะ (I want to deposit the money)
Microsoft: How do I deposit the money?
Sinitic: I want to deposit the money
สวัสดีค่ะเชคใฟ้ทีค่ะ พอดีถอนเงินไว้นานแล้วยังไม่เข้าเลยอะค่ะ (Hello, could you please check for me? I withdraw money for a long time, but do not get it.)
Microsoft: Hello, check FAI T? A long time withdrawals are not yet.
Sinitic: Hello, could you please check for me? I withdraw money for a long time, but do not get it.
ทอนคอม (Commission Bonus)
Sinitic: Commission Bonus