Let’s say you are considering adopting neural machine translation (NMT) to speed up the rendering of your global content for different locales. As you may have heard, it’s a paradigm shift in MT technology. Among the many questions that are perhaps swarming your head, we try to answer a few in this post, such as: What makes NMT work? Can it be the panacea to all your translation troubles? Can it work on its own, without any human intervention? Can it deliver the quality you’re looking for?
An NMT engine can only be so good as the data it learns from. But unlike its predecessors such as statistical and rule-based machine translation, NMT learns on its own. That is, you don’t have to teach it word by word, phrase by phrase, or by grammatical rules. It learns from correlations in the data, employing methods such as transfer learning.
Alon Lavie, VP of Language Technologies at Unbabel, said on a podcast that we don’t have much control over what these systems are learning. Hence, it can learn un-intended stuff, too. He underscored the need to retain strong control over the translation process, particularly terminology. That makes it a no-brainer to involve professional linguists in the training process.
NMT engines do better with domain- and language-specific training. For instance, ecommerce engines trained in certain language combinations may do well in those domains and languages over generic engines, but may not perform the same way in other domains and languages. Hence it’s important to perfectly customize and continually fine-tune engines.
Data modeling improves NMT quality. This refers to the process of analysing the data and closely studying the impact of changes made to the data. Do the changes introduce a bias in the data? Is the NMT engine producing the desired quality after being trained on this data? When it does, you have a model. Data modeling needs to be done by linguist engineers who have experience working with MT. It is a critical task as it can influence the output quality hugely.
So, how does the work of these professional linguist engineers impact the functioning of the MT engine?
- Firstly, one has to understand the limits of artificial intelligence (AI). Biases may be of many types. For instance, Forbes magazine wrote about how if one is using AI training sets to understand the ideal customer base for a game, the company may be led to believe that males in the age range of 15-34 would be the ideal targets. However, other age ranges or females may never have been marketed to by the company, hence producing an absence of data on those age or gender groups. In translation, too, the parallel corpora used for training NMT engines may have biases built in. Google Translate found that its NMT models reflected social biases such as those based on gender and led to biased translations: its engine had translated “He/she is a doctor” into the masculine form, and the Turkish equivalent of “He/she is a nurse” into the feminine form. Linguist engineers, from their experience in linguistics as well as language technology, are able to identify and correct such biases.
- Secondly, they also look to disambiguate. That is, they train the system on how to differentiate, for example, between brand names and their commonly understood meanings. Many a time the machine is not able to tell between different accents, as is the case with the many variants of Spanish. Here, too, linguist engineers are needed to step in and guide the engine to make language- and culture-specific decisions.
- When NMT experts manipulate the engine either to correct bias or any other element, they are able to measure the results of such manipulations. In the absence of such measurement, again, not only may erroneous translations be produced, but it will also be difficult how to trace the root of these errors. This means that working with NMT requires a virtuous circle of testing and improvements, and testing the impact of those improvements, and so on.
- Linguist engineers understand your domain and how different language combinations work. They can fine-tune the engine for domain relevance, as well as fluency and coherence. This elevates the quality of output from that of a generic engine.
In sum, how good your NMT engine output is will depend on (a) the data that you feed it and (b) the professionals you hire to harness this data.
It’d be simplistic to think that once you switch on NMT, translation will happen without human intervention. True, NMT and other language technologies bring unparalleled capabilities to translation, but it still has to be human-enhanced to be of any value to us.