Shirley Dong examines the development and problems of machine translation in the patent field

With the rapid development of AI technology, various industries are trying to incorporate AI technology in a variety of different fields, such as automatic driving, image recognition, painting, composition, and even technological innovation using AI. So, how well is AI incorporated into the patent field? As a matter of fact, some good AI applications have been developed in patent search, patent translation, patent drafting and other aspects.

This blog mainly explores AI patent translation.

Which AI technology does machine translation use?

Machine translation (MT) mainly uses neural network technology. In the process of translation model training, the accurate input and output are given to a multi-layer neural network, and the multi-layer neural network will find some coefficients in hidden layer nodes, so as to determine a translation model that can automatically output translation results.

From the above principles, it is easy to see that the quality of input and output plays an important role in the quality of the translation model. In the patent translation field, the input is the original text, and the output is the accurate translation.

How to measure the quality of MT?

At present, the popular automatic evaluation method of MT is the BLEU (Bilingual evaluation understudy, BLEU) algorithm, developed by IBM. The BLEU algorithm uses an n-gram matching rule, that is, the proportion of N groups of words between the MT result and accurate translation. For example, if the MT result is “it is a nice day” today and the accurate translation is “today is a nice day”, there are 6 words in the MT, and 5 words hit the accurate translation, therefore the BLEU value under the 1-gram rule is 5 / 6 = 0.83.

The idea of the BLEU algorithm is that the more similar the MT result is to the accurate translation, the higher the BLEU value, which means the better the quality of the MT. 

Development of AI patent translation

It is not difficult to find a variety of AI translation tools covering different languages in various patent information annual conferences. These AI translation tools are easy to use, and some of them are good tools for patent translation.

This article takes the Chinese-English translation as an example to compare the Chinese-English translation of Google, Baidu, WIPO and PremiWord. In the patent application translation field, the highest average BLEU value for MT is 0.55. In the Office Action translation field, the highest average BLEU value for MT is 0.65. In other words, the similarity between the MT and accurate translation is more than 55% (see Annex 1).[1]

In addition to the comparison of BLEU values, we take a claim as an example to show the results of MT intuitively (see Annex 2)。

From the examples, we can see the development and problems of MT in the patent field:

1. Although the MT can express basic meaning of the original text, it still needs to be improved to achieve translation acceptability;

2. Translation errors

MT in the patent field has made progress. Although there are still some errors, the optimal translation result of MT is getting closer to the accurate translation. In addition, the legal issues of the translation still need to be dealt with professionally by the patent agent according to the context and local legal requirements.

It is worth mentioning that another advantage of MT is that it greatly improves the processing efficiency. On average, 10000 words can be processed in one minute.

What documents can MT be used on in the whole patent process

In view of the current development of MT, It is believed that MT can at least be used on the following:

Although the BLEU value of MT still needs to be improved, with the continuous improvement of AI technology, MT will play a role in patent filing, office action responding, litigation and more.
We look forward to the opportunities and challenges brought by technological change to the patent industry.

Next steps

  • Consider joining FICPI’s Study & Work Committee on Software, High-Tech and Computer-Related Issues (CET 6).

How FICPI makes IP attorneys more effective

Membership of FICPI ensures that insights from the FICPI community are shared with other members, driven by a strong shared interest and offering the opportunity to learn from peers in different practice specialisms and different countries and regions around the world.

_______________________________
[1] Google website:https://translate.google.cn;
Baidu website: https://fanyi.baidu.com/;
WIPO website: https://patentscope.wipo.int/translate/translate.jsf;
PremiWord website: https://www.premiword.com.cn