Offline Translator for Financial Texts


Financial service providers often have to deal with a large number of foreign-language internal documents like e-mails or process descriptions which require translation.

Methods from the areas Deep Learning / Artificial Intelligence, on the other hand, enable the high-quality translation of demanding texts. However, although online tools would be available for efficient translation of these texts, they must generally not be used for security and confidentiality reasons. This is particularly the case for texts relating to personal data – for banks as well as all other institutions and companies alike.

We have available procedures for translating financial texts in various languages that can be used offline. With that, automated translations of confidential texts are possible without security and privacy issues.

Thus, the translation of internal texts can be significantly accelerated in order to save time and resources.

Foreign-language Texts in Financial Institutions

Due to their involvement in various countries, financial service providers usually have to deal with a large number of foreign-language texts. These can be contracts and legal documents, but also internal (partly forwarded) e-mails, documentations, presentations or remarks of the supervisory authority of the respective country. Even if the company language is e.g. English, foreign language texts cannot usually be avoided.

Numerous Internet companies have recognized the great demand for ad hoc translations and offer well-known corresponding online tools. However, translations by such procedures are always carried out by external servers, often located outside the EU, which entails certain risks. Due to high confidentiality standards – and not least because of the GDPR – the translation of internal texts via online tools is often prohibited in financial institutions.

Thus, in general financial service providers currently can not take advantage of automated translation. Instead, one of the following approaches is usually chosen

  • Engaging (external or internal) translators for high-quality translations. This is costly and relatively time-consuming, which is why it is only suitable for documents such as contracts. Automated tools will not change this in the foreseeable future.
  • The employees of the departments prepare the translations manually, e.g. for presentations or process documents, for the case they know both languages. This often takes up scarce resources at the expense of other activities and can be demotivating for many employees. Automated translation can significantly speed up processes in this case.
  • Employees who do not understand a language ask informally colleagues who do so, e.g. in the case of e-mails. This places considerable demands on the latter, which is why automated tools can also help here.

Deep Learning-based Translators

The enormous increase in the amount of data and computing power available in recent years has led to a considerable expansion of the analysis options to include even previously unquantifiable data with the help of advanced data science methods.

Particularly great progress was made in the field of deep learning, in which multi-layered (“deep”) neural networks are “trained” to analyze complex data such as texts, images or films. In the meantime, an entire “zoo” of complex neural networks for a wide variety of tasks has emerged.

Machine translation usually uses the encoder-decoder architecture. Each word of the source language is transformed to an abstract representation, taking the context into account. This transformation is then mapped to the target language.

The Transformer model – initially developed by a Google research group – uses a “self-attention“ mechanism which iteratively assigns a context-dependent score to each representation according to the scores of the other words. This state-of-the-art model is relative fast to train and provides high-quality results.

Preparing a translation model (like the Transformer model) requires several steps that vary in complexity and are described in the following

  1. The first step is to design the neural network architecture itself. This step can be skipped, however, if predesigned architectures (like the Transformer model) are already available.
  2. The training data – i.e. already translated texts of the start and target language – must be obtained and prepared in a suitable form. This step is generally the most labor-intensive as it requires the preparation of huge text amount of several GB.
  3. The model itself must be trained. Though this step runs automatically, the training is very computing-intensive; training on usual CPUs can last weeks; NVIDIA GPUs can accelerate training by a factor of 5. The model must be also validated in-sample and out-of-sample. Specific metrics like the BLEU score (bilingual evaluation understudy) can be used for that.
  4. When the model is ready, it can be called by a stand-alone program without going through the previous steps.

Our Offline Translator for Financial Texts

Our solution is written in the programming language Python and uses free libraries and models. The setup for training is not required for the use of the final translation models since the translation itself is operated by separate routines. Concerning the hardware requirements, modern laptops or desktops are sufficient, the programs run under Windows as well as under Linux. All routines run offline and need no Internet once installed.

As training data, we used huge amounts of text corpora (several GB), which are partially provided by the EU Parliament and encompass everyday as well as legal/political language. These texts have been enriched with bank- and insurance specific regulations like CRR or Solvency II in order to build a subject-specific translation tool.

We performed the training during several weeks on our own infrastructure. By using the state-of-the-art Transformer model and huge amounts of texts, we reached a very high BLEU score of over 40%. This comes close to the best achieved results – and we expect a further increase by extending the training using modern GPUs. (BLEU score can reach from 0% to 100%, where 100% is practically never reached, since it would require translations identical to reference translations.)

Sample translation from English to German
Sample translation from English to German

Our model can be operated by Python programs which can be – as the model itself – be locally installed on laptops and desktops or on intranet-accessible servers. The model itself occupies some GB on the hard drive and needs some GB RAM. Although the training is very time-consuming, the translation itself is quite fast and takes some seconds per sentence; e.g. Solvency II can be translated in less than 20 minutes on a GPU.

Each model is one-direction (e.g. English → German). New models can be trained and used for every language combination in the same way, hoever. Currently, we support the languages English, German, Italian, and French.

Our Offer – AI as a Module


  • Our translator provides high-quality translations of finance-related texts.
  • Currently, the languages English, German, Italian, and French are supported; further languages are in progress.
  • Our translator is provided as AIaaM solution and consists of a documented pre-trained neural network.
  • The solution runs in Python and requires appropriate standard libraries.
  • In order to operate the translator, the neural network files can be addressed via simple Python scripts.


  • As with online tools, translations into common languages can be speeded up and insights into texts of unknown languages can be gained.
  • Since our tool runs offline, confidencial texts can be translated without security or privacy issues.
  • After transfer, the tool is at the customer and there is no dependence on web services.
  • There is no separate installation required; the files can easily be copied to a directory and accessed via Python scripts.
  • The network can be integrated flexibly in each environment without any restrictions.
  • In case of already available Machine Learning classifications (e.g. for fraud detection), the analyzes can efficiently be extended to other languages by connecting the translation module upstream.

Depending on customer requirements, flexible design and training on demand possible.


Dr. Dimitrios Geromichalos
Founder / CEO
RiskDataScience UG (haftungsbeschränkt)
Theresienhöhe 28, 80339 München
Telefon: +4989244407277, Fax: +4989244407001
Twitter: @riskdatascience