AI Commoditization: Is the Machine Translation Industry Concerned?

Sign up for our newsletter on globalization and localization matters.

AI commoditization is now a confirmed fact for several AI domains. From a general perspective, the general public is constantly looking for freeware versions for their software needs, so there is a high expectation of freeware. At the same time, GAFA offer free services to monetize their users’ data.

Regarding the AI implementation part, the following parts of AI have been significantly commoditized because of low technological effort for the customer to access the technology, good quality result, and low price:

• Speech recognition with solutions such as Apple Siri, Google Assistance, and Amazon Alexa.
• Chatbots with platforms such as Chatfuel or Manychat, that are offering chat session management at $0.03 per session.
• Image recognition with systems such as Google Photos and Facebook’s DeepFace, that use convolutional neural networks to identify and classify objects, people, and places in images.

What about machine translation?

When it comes to machine translation (MT), there are different perspectives:

- Ease of use of some professional versions of MT software, where for less than $10 per user you have access to:

• An unlimited number of translated words.
• A personalized glossary and a plug-in interface: this means that MT can be integrated inside CRMs, CAT Tools, TMSs, and so on.

From the quality perspective, several public benchmarks have confirmed that MT can deliver very good results. By adapting the glossary and the post-editing, organizations can answer the specific vertical content or tone of voice. However, special attention should be paid to MT performance when it comes to the Legal, Financial, and Healthcare sectors because performance is not the same for all MT systems.

From language perspective, the number of proposed language combinations is increasing rapidly because the capacity of the Deep Learning system (where 2 AI systems operate: a regular MT system and a system that continuously trains the MT based on the data sets) allows for a faster expansion to new languages.

From an economical perspective, if we look at a specific translation segment, we see that the price by the translated word has evolved as follows:

Other changes are taking place:

• -Large platforms such as video broadcasters (YouTube), meeting platforms (Teams, GoogleMeet, etc), CRM editors (Salesforce, Zendesk, etc), are adding subtitling and/or translation services to their offerings.
• BPO players, such Webhelp, Teleperformance, and Armatis, which are usually far from the MT world, have developed their own integration of MT for their customers. They offer translation services at a very low cost to increase the interest of their customers to use their services. They also use their offshore agents to train the system, which can be a very efficient way to converge to a high level of translation quality.
• Customers are also creating disruption by themselves by using MT for their own translation and integrating their glossaries and managing post-editing with their in-country management.

The answer to the commoditization of MT seems to be a positive one, but if we are looking at how MT is used, it is rarely used as a standalone application. More often, it is deeply integrated into complex environments (TMS, CAT Tool, CRM, etc) and processes, where productivity, quality, and reliability are very important.

Also, if we consider the different steps of the translation process (MT, CAT Tool/TMS, post-editing), the cost of MT per word can be very low. But according to the quality output of MT, the cost of post-editing per word can be very high to close all the quality gaps. The commoditization of MT is clearly limited by the integration into its environment.

Looking at the value chain of an MT translation with post-editing can be very instructive in predicting the next steps in the industry.

Observing these data we can form a new, broader vision of the situation:

• The reduction in the cost per word of MT is due to technological progress, but also to the ability to adapt the translation environment for a limited price to what needs to be translated (glossary, memories management, training of data set , and reuse of already existing trained environment, etc).
   
• The CAT/TMS combination can help to increase productivity and the quality of the translation as well as to select the most appropriate MT for a purpose, the cost per word can vary a lot according to the pricing model (fixed, variable) and the volumes;-Post-editing progresses are a combination of TMS/CAT effects for productivity with all the helpers and MT progresses reduce the number of errors and consequently the time to correct. ChatGPT is not included in these costs, but it is expected that its integration will further reduce the price.

The situation is also creating new requests, for example:

• QE (Quality Estimation): As the usage and volume of words translated by MT increase, it is important to understand which segments translated by MT require a more accurate post-editing.
   
• Call to Action based on QE results: When QE is not at the level expected. For example, if a MT has integrated MQM (Multilingual Quality Metric) measurement inside its delivery, it will deliver the translated segment and the MQM translation score (from 0 to 100). Developing the example, if the QE is below 95, it could be extremely insightful to provide a call to action proposition in order to have the problem not reappear. This could be enhancing the glossary with a specific word, integrating new translations in the memory, or indicating the presence of irony that needs a specific translation. These could be suggestions proposed by a specific QE Call to Action system. The impact on the cost could be on the post-editing part of the process reducing the global time of review (segments with high QE scores do not need deep review) and after some time of adaptation reducing the post-editing time.
   
• Segment or traffic sorting, use the capacity of AI to select inside the segments or the traffic flow to translate which ones need to be, for example, translated by human translators. For example, in the case of a large hotel chain that needs to treat everyday thousands of customer messages in more than 35 languages by translating them in order to have them treated in a 24/7 shared customer services center with only English-speaking agents. It could be worthwhile to identify among all the emails received an email in German of a customer request complaining about food poisoning in their Berlin hotel, mentioning several days of sick live, and joining a medical certificate. This could be done by a front-end Artificial Intelligence reviewer system of the emails that will tag this specific German email as a subject with a potential impact on the brand and legal implications. According to the customer choice, this email with legal implications might not go inside the MT flow for translation and further treatment by the English-speaking shared customer service team but in a German native customer service queue for specific premium treatment. Also, with similar customer decisions, the emails with brand reputation implications and a low MT QE score might not go too inside the MT flow. Integrating this triage inside a Salesforce queue management system can be very interesting and help to remove barriers to MT adoption.

The control of the entire process line with all implications is providing in this environment a key competitive advantage starting with a deep knowledge of what is the type of document to translate, the expected quality, and the expected revision. In recent times, we have seen emerging companies specializing in one part of the Localization market (Website translation, Subtitling, etc) and proposing only one very lean service line at a very competitive price and very easy-to-use model.

The control of the process is very important because this deep knowledge allows us to anticipate all the integration requests that can come from the customers and either develop them or propose adapted plugin strategies. Integration of QE can also bring an immediate advantage to these companies because this can facilitate customer adoption.

Agility and strong engineering expertise associated with an accurate vision are key in a fast-moving market because. All these elements generate inflating requests for integration and adaptations which are also opportunities.

Do you want to contribute with an article, a blog post or a webinar?

We’re always on the lookout for informative, useful and well-researched content relative to our industry.

Write to us.