Machine Translation: Neural or Neutral?cotuong
The machine translation (MT) technology landscape has experienced profound and seismic shifts since 2016. Announcements from tech giants Google, Microsoft and Facebook regarding the dramatic quality improvements of neural machine translation (NMT) and coverage by mass media catapulted NMT onto everyone’s radar. Almost overnight, NMT became the “holy grail” of MT. This white paper provides an overview of the technologies; attempts to explore the notion that NMT should be put into production immediately; and argues for a measured approach when selecting between neural, statistical (SMT) and rules-based (RBMT) engines, as each has a role to play.
Unlike RBMT engines based on explicit grammatical rules or SMT engines which build probabilistic models based on language models and phrase tables, NMT models use neural networks where words are represented by vectors and have semantic relationships. NMT engines employ bidirectional recurrent neural networks (RNNs) or convolutional neural networks (CNNs) first to encode source text and subsequently to decode or predict translations. NMT has the potential to dramatically improve translation quality, especially in languages where word order deviates between source and target such as Japanese. SMT engines still tend to do very well for longer sentences, Romance languages and have robust lexical coverage when trained accordingly. RBMT engines are still valuable for normalization, translation from one adjacent locale to another — such as Spanish (Spain) to Spanish (neutral) — and development of rare, long-tail language translation directions.
When considering NMT, it is highly recommended to consider certain factors and assign priorities to each: Infrastructure and cost: To be practical, NMT engine training requires GPUs (not CPUs). Thus, the hardware is far more expensive than SMT or RBMT engines. This cost will likely be reflected in licensing fees. Also, in deployments of any new MT technology, it is important to consider the cost of connectors as well as customization for tag handling.
Training and maintenance: Due to the intensive computing resources to build NMT models, it takes longer to build these engines. Enforcing specific patterns from linguistic feedback during retraining is more difficult due to a fundamentally different technology; it’s not a matter of modifying phrase tables or language models as with SMT or rules/ dictionaries as with RBMT.
Quality: NMT output is remarkably more fluent than SMT and RBMT, and particularly so for Asian languages, where the latter two have always been relatively weak. However, this fluency does not guarantee accuracy. The cognitive load can be higher for a post-editor to review the source and suggested target.
Data: As with SMT, your engines are only as good as your own data. In tests conducted by Welocalize, we found that NMT systems typically require a lot more training data than comparable SMT or RBMT systems.
Players: Much like SMT, the field is wide open, ranging from the usual suspects, such as Google and Microsoft Translator, which may soon be customizable, to open source initiatives, such as OpenNMT and Nematus, to boutique MT providers and DIY platform providers where you bring your data and train your own engines.
In some ways, the advent of NMT is like the rise of robots for automating daily tasks. Both technologies are extremely promising, yet some tasks are still better suited to current technologies. NMT engines can be implemented, but only after careful considerations of initial and ongoing costs for engine training, technology – connectors, translation management systems (TMS) and computer-aided translation (CAT) tools – as well as unambiguous evidence of better quality relative to SMT or RBMT for specific language pairs, improved linguist productivity and lower translation costs .