Since 2011, Patent Translate has been overcoming language barriers and opening up access to the world’s patent documentation. And recently it has improved even further, thanks to a technology called “neural machine translation”.
Marking the EPO as a pioneer in machine translation of patents from the moment of its launch, Patent Translate offers translation between English, French or German and all the official languages of the EPO’s member states, as well as Chinese, Japanese, Korean and Russian.
Usage statistics show that Patent Translate is a global service. The 15 000 translation requests each day come from across the EPO’s member states and beyond, with a large proportion originating from India, Japan, Russia and the United States. This number is a clear illustration of the need the EPO was responding to when it launched Patent Translate.
Global patent data is multilingual and users demand seamless and immediate access to it. Patent Translate has clearly moved us close to that goal. EPO data at the core This remarkable achievement has, in part, been due to the sheer amount of data the EPO can provide. A prerequisite for good
machine translation is the existence of high-quality human translations between the two languages concerned.
Example of machine translation with Google’s old and new translation engines:
| Original text | Old translation engine | New translation engine |
| Die Erfindung bezieht sich auf ein Verfahren zum Konfektionieren eines Jumperschlauch-Kabels, welches aus einem flexiblen Schutzschlauch und darin geführten Kabelleitern besteht, bei welchem zumindest ein Teil des Schutzschlauches in seiner Längsrichtung zur Schaffung einer ausreichenden Leiterüberlänge gestaucht wird, sowie auf eine Vorrichtung zum Konfektionieren eines Jumperschlauch- Kabels, welches aus einem flexiblen Schutzschlauch und darin geführten Kabelleitern besteht. |
The invention relates to a method of building a jumper tube cable which comprises a flexible protective tube and guided therein cable conductors, wherein at least a portion of the protective tube is compressed in its longitudinal direction to provide sufficient head excess length, as well as to an apparatus for assembling a Jumper schlauch- cable that consists of a flexible protective tube and guided therein cable conductors. |
The invention relates to a method for assembling a jump tube cable consisting of a flexible protective hose and cable conductors guided therein, at least one part of the protective hose being compressed in its longitudinal direction in order to provide a sufficient conductor overhang, as well as a device for assembling a jump-Cable, which consists of a flexible protective hose and cable conductors. |
This solution has the power within to work against the functions of PDE5 enzyme, the chemical that was discovered to obstruct levitra vs viagra the mechanisms of cyclic guanosine monophosphate. Magnesium deficiency can affect virtually cialis generika you can look here every organ system of the body. viagra in italy This medicine is not hard or in large form. This drug is manufactured by the Ajanta Pharma in India with facility certified by the Food and Drug Administration for the treatment of erectile dysfunction, commonly known as tennis elbow, and medial cialis 5mg no prescription epicondylitis, more commonly known as golfers elbow, are conditions which describe pain and inflammation of the tendons and muscles at the elbow joint.
The computer can analyse these human translations and learn from them. To do so, it needs a lot of translations. The patents in the EPO’s databases are perfect for this purpose. The EPO was able to provide Google with the text from hundreds of thousands of patent documents in different languages. Thanks to the fact that the databases gather the documents together in patent families, it is known which texts in different languages correspond to each other.
The EPO sets stringent requirements for the data it provides to feed the translation machine, insisting that tens of thousands of human translations must be available in a particular language before it considers offering that language in Patent Translate.
Access to foreign language data Machine translation is not just replacing human translation, but is providing access to masses of data previously inaccessible to patent searchers. Back in 2013, for example, the EPO estimated that it would take 16 000 man-years to translate the Chinese patent documentation available at the time into English.
Patent Translate provides all that documentation, and the additional documents published since then, instantly in English, French and German, and to a level of quality that is perfectly adequate for readers to understand the nature of the invention. It would have been impossible to translate this information using human translators. Neural machine translation Neural machine translation (NMT) takes machine translation a step further and puts the EPO/Google co-operation at the forefront of machine translation technologies. NMT is already in use in Patent Translate for eight languages (in addition to English) – Chinese, French, German, Japanese, Korean, Portuguese, Spanish and Turkish.
Access to foreign language data Machine translation is not just replacing human translation, but is providing access to masses of data previously inaccessible to patent searchers. Back in 2013, for example, the EPO estimated that it would take 16 000 man-years to translate the Chinese patent documentation available at the time into English.
Patent Translate provides all that documentation, and the additional documents published since then, instantly in English, French and German, and to a level of quality that is perfectly adequate for readers to understand the nature of the invention. It would have been impossible to translate this information using human translators. Neural machine translation Neural machine translation (NMT) takes machine translation a step further and puts the EPO/Google co-operation at the forefront of machine translation technologies. NMT is already in use in Patent Translate for eight languages (in addition to English) – Chinese, French, German, Japanese, Korean, Portuguese, Spanish and Turkish.
The conclusion that the NMT method leads to better results is evident to anyone who uses Patent Translate with these languages. The quality is not just anecdotal, however, since the EPO and Google also use scientific methods to assess how good machine translations are. There are two measures for patent translation quality: accuracy and fluency. In both of these areas, the new NMT solution is producing significant improvements.
Google’s latest development in machine translation is called Attention, which lies behind the recent advances in Patent Translate. With Attention, the machine finds the correct translation of words and phrases by trial and error, over and over again, and establishes artificial neural pathways closely resembling those in human brains. As the number of neural pathways increases, a neural network forms.
Since this happens via the language pairs of translated patents and patent-related documents, Patent Translate is very precise. It recognises styles of writing based on context and diction and is able to make accurate translations accordingly, down to appropriate nuances in style. Attention is a system in which programmers are “making the machine understand how to focus its attention on specific pieces of the input. You can visually see where the attention is given,” says Ian Wetherbee, a senior software engineer at Google. The Attention system gives numerical feedback that indicates proximity to the most correct translation. If the system presents an unfavourable word choice, the programmers are able to locate the precise point at which the machine made an error, and correct it.
Programmers no longer have to tell the machine what to look for, because the system finds patterns based on syntax from examples in the data. In this way, the machine learns to make choices independently of a human. The system accesses words and their definitions, and sentence structures. It then applies meaning, thereby learning to choose a correct translation.
Neural networks even mean that computers can now use what they learn from one language pair to improve the translations of another language pair. With the vast quantity of data, the machine begins to predict how the language is structured, how words are used, and how to develop its own sentences in the other language, based on the relationships in previously seen language pairs. Each network can learn a global model and arrange its own internal language so that it can tackle language pairs that it hasn’t been trained for. “It can create this universal representation of language, so as long as it had enough data from one language, it could learn how to represent that in the universal model and end up translating into any number of output languages,” says Wetherbee.
With its NMT solution and with the help of EPO data, Google has taken automated translations to a new level, and several more languages are in the pipeline.